Benchmark run catalog

2026-11-03 RTX5090 candidate campaign

Configuration	Measured GPU / path	PRO relationship	Evidence labels	Decision labels	Evidence
Private candidate on Ignis 0.8.0,32K budget pair and256K vision profile	One RTX5090; direct C1	Not measured; no extrapolation	quality, functional, bounded capacity	challenger, no-promotion	Dossier · Finding · Ignis note
Private QAT/NInfer text160K and failed vision sibling	One RTX5090; direct C1	Not measured; no extrapolation	quality, functional; compatibility-only failure	challenger / rejected, no-promotion	Dossier · Finding
Private coding/NInfer128K fresh recheck and earlier distinct process	One RTX5090; direct C1	Not measured; no extrapolation	quality, functional; failed outer context job retained	challenger, no-promotion	Dossier · Finding
Private historical and restored160K vision references	One RTX5090; direct C1	Not measured; no extrapolation	quality, functional; historical-invalid (initial identity provenance incomplete)	no-promotion	Dossier · Finding

Selected candidate configurations passed123/123 coding checks; the restored reference114/123. These are seven task types repeated three times with different loaded policies/capabilities, not a complete serving comparison or universal ranking.

2026-11-01 Swift TP1 GPU0 qualification accepted

Configuration	Measured hardware / path	Outcome and limits	Evidence
Swift 1.5 Qwen3.8 27B BF16 with FP8 KV, TP1, 262,144 configured context and MTP3; selected C2 interactive operating point	One RTX PRO 6000 Blackwell Max-Q; final fresh capacity population is 10 cells / 72 requests, separate from the historical 60-cell matched comparison	Protocol and bounded qualification gates passed; no current deployment or client assignment claim. C2 32K median decode 13.43 tok/s is below its 15 tok/s soft objective while first-output p95 is 16.69 s under the 30 s ceiling; C1 is latency-first. C16 is queued requested load, not 16 active requests. SWE resolved 4/5; strict formatting failed 4/4. Operational records are private.	Dossier · Finding

2026-11-01 cache-layout and DFlash scouts

Configuration	Measured hardware / path	Outcome and limits	Evidence
Private target; original MTP3 versus internal 1M cache layout, then exact original restoration	Two RTX PRO 6000 Blackwell Max-Q, TP2/DCP2; direct online ~6K/~25K, requested C1/C4	144/144 timed requests; no 10% E2E win against both controls; +38.60% reported cache slots, ~25K/C4 7.92–9.25% slower. Natural output, absent canaries; no full-context or quality qualification. Original restored, no-promotion.	Dossier · Finding
Same private target; original MTP3 versus DFlash2 K3, then exact original restoration	Two RTX PRO 6000 Blackwell Max-Q, TP2/DCP2; direct online ~6K/~25K, requested C1/C4	144/144 timed requests; locally compatible in this screen, no 10% E2E win against both controls; −40.63% reported cache slots. No controlled output, canaries or matched no-speculation arm. Original restored, no-promotion.	Dossier · Finding

2026-10-26 [operator-private] H3 MCP functional smoke

Configuration	Measured hardware / path	Outcome and limits	Evidence

2026-10-26 Qwen3.8 Flash Next three-recipe scout — final, no-promotion

Configuration	Measured hardware / path	Outcome and limits	Evidence
Distinct pinned Qwen3.8 Flash Next recipes: Ignis 0.8.0, vLLM r2, and Flare r2; serialized C1 agentic/Aider/SWE/context scouts	One RTX PRO 6000 Blackwell Max-Q per run; single-request workload. The second card was not used for performance measurements.	Agentic 6/6 and context 9/9 passed on each recipe; context prompts reached 114,588 actual tokens. Official SWE resolved 2/5, 3/5, and 0/5, respectively; the denominator is five attempted tasks per lane, with empty patches ungraded. Flare attempted five tasks but only Django produced a nonempty graded submission. Aider outcomes were mixed and include ungraded budget stops. Shared-router whitespace handling makes historical coding attribution uncertain. vLLM C1 was 0/8 performance-eligible; Flare C1 ended 0/8 performance-eligible, with native native_control_mismatch preserved; exact prior evaluation routes/output caps and GLM incumbent readmission passed local checks; the router uses reviewed PR #704 overlay, not the original image. Final no-promotion; no engine ranking, speed, full-window, or concurrency claim.	Qwen dossier · Finding · Evidence index

2026-10-26 Ignis 0.8 mandatory-tool qualification

Configuration	Measured hardware / path	Outcome and limits	Evidence
Pinned Qwen3.8 Flash Next Trellis on Ignis 0.8.0; TP1, 131,072 configured context, eight lanes, MTP3; phase-one auto, phase-two required with normal thinking	One RTX PRO 6000 Blackwell Max-Q through the direct OpenAI-compatible endpoint; second equal card protected at 0 MiB candidate allocation	Original unseeded auto batch remains 3/4. Same-runtime phase two passed 3/3 mechanism cells, selected full preflight 5/5 checks and 9/9 observations, and seed-18 tools 4/4. Sixteen terminal engine events; HTTP 200 retained for 12 captured tool requests. Exact operational restoration passed after both phases. Functional request-policy evidence only; no coding, capacity, performance, client, or promotion claim.	Qwen dossier · Finding · Summary · Evidence index

2026-10-26 GLM reasoning-close GPU marker trace

Configuration	Measured hardware / path	Outcome and limits	Evidence
GLM FP4 with pinned split-top-p patch and bounded marker trace; TP2/DCP2, FP8 KV, 327,680 configured context/C4, MTP3; one C1 probe	One RTX PRO 6000 Blackwell Max-Q for actual-CUDA startup gates; dual-Max-Q model request through route-derived direct relay	Preweight selftest and 272-row sampling gate passed. The one exact-answer response had one close at index 64 with matching budget-store events on both ranks, but the trace reported 76 tokens versus 75 from the API and the frozen analyzer returned unknown. No second close reproduced. Exact operational restoration passed; strict JSON failed before and after, with no established regression. No promotion, resolution, capacity or performance claim; Ignis held.	GLM dossier · Finding · Summary · Evidence index

2026-10-25 prospective sampling upstream-tolerance trial

Configuration	Measured hardware / path	Outcome and limits	Evidence
GLM FP4, pinned split-top-p patch; native-count gate; TP2/DCP2, FP8 KV, 327,680 configured context/C4, MTP3; one C1 probe	One RTX PRO 6000 Blackwell Max-Q for synthetic kernels; dual-Max-Q model request through route-derived direct relay	The prospective upstream count gate passed all eight forced cases and 272 ordinary rows. One model request avoided the earlier token-0 loop but emitted closing markers at positions 64 and 126 and failed strict output. Exact operational restoration passed; matched JSON smokes failed before and after, with no established regression. No promotion; Ignis held. No performance comparison.	GLM dossier · Finding · Summary · Evidence index

2026-10-25 split-top-p sampling-reference diagnostic

Configuration	Measured hardware and layer	Outcome	Evidence
GLM FP4 candidate image with exact upstream split-top-p patch and independent binary64 numerical reference; model weights not loaded	One RTX PRO 6000 Blackwell Max-Q, synthetic CUDA kernel check; two cards reserved for the unrun TP2 candidate	8/8 forced cases passed. Stock/patch masks were identical on 271/272 ordinary rows; the patch kept one extra token and failed the strict local no-worse gate. Zero candidate requests. Exact operational restoration passed; the strict JSON smoke failed, leaving functional acceptance incomplete. No promotion; Ignis held.	GLM dossier · Finding · Summary · Evidence index

2026-10-25 pinned sampling-fix kernel trial

Configuration	Measured hardware and layer	Outcome	Evidence
GLM FP4 candidate image with exact vLLM split-top-p backport; model weights not loaded	One RTX PRO 6000 Blackwell Max-Q, synthetic CUDA device-0 kernel check; two cards reserved for planned TP2	Ordinary parity failed after partial forced-case success; zero candidate model requests. Functional, no promotion, Ignis held; end-to-end fix unproven.	GLM dossier · Finding · Authoritative counts · Evidence index

2026-10-25 isolated reasoning-boundary diagnostic

Configuration / scope	Hardware and evidence	Outcome / boundary	Links
GLM NVIDIA FP4 incumbent; C1, numeric budget64/output512	Dual RTX PRO 6000 Blackwell Max-Q; functional diagnostic	Failed boundary transition; no-promotion, Ignis held, runtime unchanged. See canonical finding for measured token pattern and limitations.	GLM dossier · Finding

2026-10-25 GLM Go/Rust completion-cap diagnostics

Configuration / scope	Hardware and evidence	Outcome / boundary	Links
GLM-5.3-Flash; pinned Aider/Polyglot Go and Rust tasks with 16K/32K completion caps	Dual RTX PRO 6000 Blackwell Max-Q measured; complete engine/GPU/host/cgroup telemetry and verified closure for all six submitted jobs	Six completed measurements: one passing/five failing task grades; two supported cells unstarted, four reasoning controls blocked. Raw grades remain separate from lifecycle and measurement eligibility. No causal cap benefit, Ignis run or promotion.	GLM dossier · Finding · Evidence index

Closure through the final postcheck took 26.73 minutes within the 45-minute operator limit. Three additional Aider request starts lack captured responses; their usage remains unknown. See the authoritative summary.

2026-10-25 R12 baseline and R13 admission refusal

Configuration / population	Hardware and evidence	Outcome / boundary	Links
GLM NVIDIA FP4, vLLM TP2; R12 incumbent	2x RTX PRO 6000 Blackwell Max-Q; C1/C4 native tools and contained coding tests	48 submitted: 44 complete measurements/grades (38 pass, six fail), four cancelled unknowns. Released batch closure verified; request overlap unknown. Planned 48-cell candidate arm unrun.	GLM · Finding
Qwen3.8 Flash Next Trellis, Ignis 0.8; separate R13	One RTX PRO 6000 configured, no candidate container created	0/48 submitted; 49 GiB RAM plus 16 GiB host reserve refused before creation, exact available memory unknown. No candidate quality, timing or capacity result.	Qwen · Finding

Exact incumbent restoration and explicit minimal-thinking Pi acceptance passed; historical thinking-off failure remains open. No matched comparison or promotion. Authoritative summary · Evidence index · Manifest.

2026-10-24–25 evaluation-platform pilot

The following rows retain the historical R5–R10 cutoff. The successor above records later evidence without changing those outcomes.

Configuration / population	Hardware and evidence	Outcome / boundary	Links
GLM NVIDIA FP4, vLLM TP2; R8 incumbent arm	2x RTX PRO 6000 Blackwell Max-Q executed native tools and contained coding tests	42 complete measurements/grades: 32 pass, ten fail; two maximum-call failures ungraded, two cancelled, two still cancelling; final interval unknown. Separate mixed-C4 canary closed four complete measurements with three passes/one valid task failure. Earlier attempts remain separate.	GLM · Finding
Qwen3.8 Flash Next Trellis, Ignis 0.8; R9/R10	One RTX PRO 6000 configured; R10 startup only, no generation measurements	R9 host-reserve refusal; R10 monitoring identity-adapter block. Each remains 0/48 submitted. Exact incumbent service restored; strict Pi final output failed. Fresh post-fix candidate campaign and accepted Pi oracle still required; no paired comparison or promotion.	Qwen · Finding

2026-10-21–23 Flare engine investigations

Date	Scope	Decision	Evidence
2026-10-21	Initial candidate scout and configuration optimization	no-promotion; different checkpoints and incomplete replacement gates prevent an engine-only ranking	GLM dossier · Scout · optimization
2026-10-21–22	Inference and cache-row diagnostics	no-promotion; retained counters and replay observations are diagnostic, not a performance qualification	GLM dossier · Inference · cache
2026-10-22–23	Allocation, protected-regression, and upstream compatibility controls	no-promotion; numerical/protected gates remain held and the incumbent was restored	GLM dossier · Allocation · state · refresh

2026-10-19 Ignis 0.7.2

Configuration	Hardware	Result	Decision	Evidence
Qwen3.8 Flash Next, Ignis0.7.2, 131K configured, MTP3, eight lanes	1x RTX PRO6000 Blackwell Max-Q	C1 5/5; short C8 tools8/8; SWE incomplete: one graded unresolved, four empty ungraded	no-promotion; incumbent restored; no performance comparison	Qwen dossier · Finding

2026-10-17 GLM NVIDIA FP4 primary promotion follow-up

Date	Configuration	Measurement	Hardware	Result	Decision / limitation	Evidence
2026-10-17	nvidia/GLM-5.3-Flash-FP4-Local 76348ddb2be2e35e39ecfc00e75dc07703f4f3d5, TP2/327,680/C4/MTP3	direct and authenticated routed acceptance; fresh single-arm client C1/C2/C4 capacity acceptance at approximately 31.6K actual prompts	2x RTX PRO 6000 Blackwell Max-Q	direct 9 protocols; routed preflight 8/8, image 18/18, coding/tool continuation 1/1 across five requests; capacity 12 scored requests with no observed OOM, preemption, or refusal; five fresh Pi process combinations, two primary Hermes CLI profiles, Companion Node/Mac OpenClaw Gateway and WebUI catalog checks passed	operator-selected best current fit; human-approved current primary mapping; earlier A/B remains canonical. Client C4 sampled maxima were engine active 2, queued 3, occupancy 4 at separate times, not four simultaneous decodes. NativeWorkflows is recovery_required with failed progress and no retry; Earlier evaluator credential and lost-capture failures remain retained; later bounded native checks passed. Pi Web, existing sessions, running Hermes Gateway and native catalog/session propagation remain unaccepted. No SWE, endurance, p99, MTP-off, full-window simultaneous-capacity, fleet-convergence, or public-web claim. EXL3 r11 is retained rollback.	GLM · finding

2026-10-17 GLM serving-performance comparison

Date	Configuration	Measurement	Hardware	Result	Decision / limitation	Evidence
2026-10-17	NVIDIA GLM-5.3-Flash-FP4-Local TP2/DCP2/MTP3 candidate; retained EXL3 r11 TP2/DCP1 no-spec control	12 matched natural/edge pairs across four actual-input strata and C1/C2/C4	2x RTX PRO 6000 Blackwell Max-Q	24 accepted cells; 84 requests per arm; all required and supplemental coverage retained	original 42-cell matrix closed_partial (one visible-output-overrun failure, 17 stop-rule-blocked); no SWE, endurance, no-speculation, or MTP-off comparison; no qualification or promotion. Exact r11 restoration passed direct/routed smokes and readmission.	GLM · finding

2026-10-16 GLM natural-output baseline

Date	Configuration	Measurement	Hardware	Result	Decision / limitation	Evidence
2026-10-16	GLM-5.3-Flash r11 natural-output baseline	12 natural timing cells plus two five-task coding repeats	2x RTX PRO 6000 Blackwell Max-Q	239/240 timing retained; R1 4/5 graded resolved; R2 0/2 graded resolved plus three step-limit failures	no-promotion; no pooled coding score and repeatability not established	GLM · finding

2026-10-16 SGLang, Qwen, and Nemotron comparison

Date	Configuration	Measurement	Hardware	Result	Decision / limitation	Evidence
2026-10-16	Original GLM baseline, GLM SGLang scout, Qwen FP4 scout, and Nemotron 3.5 Super VL NVFP4 scout	Native preflight, Oracle, context/vision where completed, strict C1, and terminal five-case SWE	2x RTX PRO 6000 Blackwell Max-Q	Original GLM C1 0/4 performance-eligible; SGLang SWE 0/5 resolved, 2/5 graded; Qwen SWE 2/5, 4/5 graded; Nemotron preflight 6/6, vision 18/18, context 2/2, Oracle 48/54, C1 0/4 eligible, SWE 2/5 resolved and 4/5 graded	no-promotion; SGLang/Qwen coding miss the 4/5 floor; original GLM and Nemotron C1 are ineligible, and Nemotron misses exact Oracle. No eligible controlled performance population.	GLM · Qwen · Nemotron · finding

2026-10-16 GLM v0.9.1 trial

Date	Model / lane	Scope	Hardware	Result	Decision	Evidence
2026-10-16	GLM-5.3-Flash v0.9.1 R2	native preflight, strict Oracle, vision, bounded context, sampled reserve, and pinned SWE	2x RTX PRO 6000 Blackwell Max-Q	R2: preflight 6/6, Oracle 54/54, vision 18/18, retrieval/tool-context at 216307/110760 actual tokens, SWE 3/5; restored original: direct/routed 6/6, Oracle 54/54, vision 18/18, matching context probes, SWE 3/5	rejected, no-promotion; exact incumbent restored, zero OOM, candidates absent, primary router readmitted; both coding results below the four-case gate; no capacity, stability, DFlash, performance, causal, or model-family claim	GLM · finding

2026-10-15 replacement evaluation

Date	Model / lane	Scope	Hardware	Result	Decision	Evidence
2026-10-15	GLM control, MiMo MOPD, Mistral Medium 3.5, and DeepSeek V4.1 diagnostics	completed replacement gates; DeepSeek Stage 2 numerical, r12 dummy-path, Stage 4 real-weight, and full-role diagnostics	dual RTX PRO 6000 for GLM/MiMo/Mistral and DeepSeek TP2 diagnostics; one GPU for DeepSeek Stage 2	no candidate qualified; r2/r3 failed exact resource gates and r4 failed its first Oracle8 round after passing resources	no-promotion; GLM retained after restoration; no performance conclusion	GLM · MiMo · Mistral · DeepSeek V4.1 · finding

2026-10-13 full qualification campaign

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-10-13	GLM continuity baseline and mixed TP2 diagnostics	Official-base GLM continuity line; C1/C4 controlled-output cells; mixed TP2 no-spec/thinking-off SWE diagnostics	2x RTX PRO 6000 Blackwell Max-Q	baseline SWE 2/5 complete; actual prompt means 25,280.5/25,281.125/106,520/216,194.67 in separate C1/C4 cells; mixed lanes incomplete at 1/5	continuity/contract baseline; no-promotion; no universal-quality or p99 claim	GLM · finding
2026-10-13	Qwen Coder390 and Flash Next	Coder390 TP1 INT4/INT8; QAD TP1/TP2; EXL3 4.05 TP1 parser-corrected lane	2x RTX PRO 6000 Blackwell Max-Q	completed Coder390 medium 3/5 and INT8 control 2/5; other lanes retain protocol/quality/compatibility evidence; top-k-20 cancelled before load/query	no-promotion; no cross-lane ranking or claimed top-k improvement	Qwen 27B · Flash Next · finding
2026-10-13	DeepSeek V4 Flash 0731 EXL3	TP1 bounded protocol/quality and TP2 compatibility lane	2x RTX PRO 6000 Blackwell Max-Q	retained functional/quality observations; TP2 is compatibility-only; invalid paired-image map excluded	no-promotion; no performance, context, or client claim	DeepSeek · finding
2026-10-13	MiMo V2.6 Flash conversion/trust lane	NVFP4 TP2 text-only offload and routed trust checks	2x RTX PRO 6000 Blackwell Max-Q	corrected 2 GiB offload selection; retained policy/trust failures	unresolved candidate; no physical-fit or quality conclusion	MiMo · finding
2026-10-13	Ignis R4 protocol and terminal SWE	TP1/C1 protocol, quality, and five-task SWE diagnostic	2x RTX PRO 6000 Blackwell Max-Q	serial protocol 6/6; strict quality 12/12; C8 preflight 6/13 with seven immediate 503 refusals; separate SWE 24 engine-full retry notices all recovered; terminal SWE 2/5 resolved, 2/5 graded, three empty	no-promotion; protocol evidence is not capacity, C4, or client acceptance	Flash Next · finding

2026-10-05 Xeon model qualification campaign

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-10-06	MiMo and uncensored GLM coding-quality scout	MiMo V2.6 Flash RL 5711b268 TP2/DCP1 no-spec; Lovesenko GLM-5.3-Flash TR3 4-bpw abliterated c8f58e6a TP2/EP2/DCP1; both configured 327,680/C4/eight images	2x RTX PRO 6000 Blackwell Max-Q	MiMo preflight 25/25, agentic 15/18 then 17/18/17/18, SWE 4/5; abliterated GLM preflight 25/25, agentic 18/18, SWE 3/5; no performance or full-context result	Both not-qualified, no-promotion; exact r11 restoration verified (direct 25/25, routed 7/7, readmitted)	MiMo · GLM · finding
2026-10-05	Qwen replacement scout	Evaluated prospective primary profile	2x RTX PRO 6000 Blackwell Max-Q	frozen primary SWE 3/5; strict C4 128-word cell 0/4 performance-eligible	rejected, no-promotion; applies only to this evaluated profile	Qwen dossier · finding
2026-10-05	GLM-5.3-Flash CSF/QAD scout	Pinned QAD revision, TP2/DCP2, C4, 327,680 configured tokens, no speculation	2x RTX PRO 6000 Blackwell Max-Q	max functional 25/25; Oracle 18/18; initial SWE 4/5; fresh-reload A1 SWE 2/5 attempts, 2/4 graded (Sphinx/Xarray resolved; Django/Pylint false; Sympy ungraded at 60 steps); exact eight-image/context/offered-C4 summaries 4/4; strict C4 0/4 performance-eligible	rejected, no-promotion; A1 misses the 4/5 floor, so B1/B2/A2 did not run. No full-window, simultaneous-C4, or speed result. Exact r11 restored: direct 25/25, routed 7/7, and fresh Pi read-tool passed; routes/manifests unchanged	GLM dossier · finding

2026-09-26 Swift TP2 campaign

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-09-26	Swift 1.5 Qwen3.8 Flash Next TP2 bounded qualification	Exact NVFP4 3ff052…, vLLM 0.30.0 digest 8a69ff…, TP2/C1, GPU-resident PLE, no CPU offload; 327,680 context and requested 65,536 output allowance	2x RTX PRO 6000 Blackwell Max-Q	protocol/tools; agentic 18/18; context 12/12 at ~32.7K and ~256.9K actual prompts; images 12/12; eight images 4/4	not-qualified, no-promotion; SWE fixed-denominator 2/5 below the 4/5 floor; exact GLM incumbent restored; no performance or full-window claim	Efficient variants · finding

2026-09-23 GLM runtime investigation

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-09-23	GLM-5.3-Flash DCP1 r11 qualification	Pinned EXL3 4-bpw/v84 image, FP8 KV, TP2/EP2/DCP1, APC, batch 2,048, C4, 0.97 memory, no speculation	2x RTX PRO 6000 Blackwell Max-Q	direct/routed preflight 28/28 each; 13 client semantic paths; retained strict-format failures on Pi/Pi Web, secondary Hermes vision, and Mac OpenClaw; 201K/310K, quality 12/12, agentic 18/18, absolute SWE 4/5	selected/deployed, publication_verified; independent reviews, dashboard imports, Pages success, and HTTP/browser readback retained. 825,268 KV is 41.73% below DCP2; no speed, C8, crash-rate, four-full-327K, or fleet strict-format claim	GLM dossier · finding
2026-09-23	Runtime stability / configuration diagnosis	Same pinned EXL3 4-bpw/v84 image, FP8 KV, TP2/EP2, APC, no MTP, batch 2,048; DCP2 parent versus DCP1	2x RTX PRO 6000 Blackwell Max-Q	Parent serial passed; 175K/32K overlap crashed with two Xid31 faults; DCP1 3/3 matched overlaps passed; reported KV 1,416,244 → 825,268 tokens	functional diagnostic, challenger, no-promotion; exact baseline restored; no speed, failure-rate or full-envelope claim	GLM dossier · finding

Eight-image follow-up, 2026-09-18: GLM EXL3 r9 passed four eight-image comparison requests, one eight-image high-resolution request. Limit now eight/request; no concurrent eight-image soak.

2026-09-22 model shortlist validation

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-09-23	Qwen3.8 Flash Next NVFP4 qualification	Pinned NVFP4 TP2 candidate; medium SWE/Oracle/strict repair, separate default-xhigh diagnostics	2x RTX PRO 6000 Blackwell Max-Q	medium: SWE 4/5, Oracle7 18/18, strict C4 0/4; default-xhigh preflight 26/26, vision 4/4, 412,589-token retrieval	no-promotion; strict failure is performance-ineligible	Qwen dossier · finding
2026-09-23	GLM-5.3-Flash Spark NVFP4 qualification and incumbent control	Pinned Spark TP2 candidate; frozen SWE/Oracle/strict C4; exact incumbent restoration	2x RTX PRO 6000 Blackwell Max-Q	Spark SWE 2/5 attempted (3 submitted/graded; 2 at 60-call limit), Oracle7 16/18, strict C4 0/16; incumbent SWE 3/5, Oracle7 18/18, strict C4 0/16	retain GLM r10 APC; no route or client change	GLM dossier · finding
2026-09-22	Qwen3.8 Flash Next replacement scout	RadixArk NVFP4 7b719225, SGLang 4ccff141, TP2 with C4 configured admission, 262,144 tokens, no MTP, GPU-resident FP8 PLE, automatic BF16 KV	2x RTX PRO 6000 Blackwell Max-Q	cold load and six serial C1 direct preflight groups passed; Qwen 16/18 greedy and 15/18 native-sampling results, plus paired GLM 16/18, cannot rank broad coding because fixture/scorer defects invalidate planning comparisons; no capacity or performance cell	retain GLM; interesting but unqualified candidate; heavy qualification blocked by harness validity, no-promotion	Qwen dossier · finding

2026-09-21 MiMo replacement qualification

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-09-21	MiMo v2.6 Flash replacement trial	Pinned MiMo v2.6 Flash RL, SGLang v0.5.20, mixed MXFP4/FP8, TP2/C1 reconstruction	2x RTX PRO 6000 Blackwell Max-Q	V3 transport health and load passed at 81.41 GiB/rank, but both bounded correctness probes returned zero visible content; no capacity/throughput claim; GLM C4 100/100 eligible; SWE official 0/1 is confounded	retain GLM, no-promotion; final restoration verified	MiMo dossier · finding

2026-09-19 APC promotion

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-09-19	Huihui Qwen3.8 final direct-I/O 160K context envelope	Pinned Huihui artifact, NInfer 70434721, baked image 22fa89d5, binary 82ec9634; MTP3/INT8 KV, 163,840 context/C1, 8,192 visible-output cap	RTX 5090	startup; direct preflight 8/8 including 144,425-token needle and 147,941-token tool; repeated exact retrieval 9/9 at 150,058/150,144/150,124 prompt tokens	qualified 160K/C1 reference; 64K remains a historical comparison; retrieval latency mixes cold/partial-cache and warm attempts	Qwen dossier / context-envelope finding
2026-09-19	GLM-5.3-Flash r10 APC	Same checkpoint/v84 image; EXL3 4-bpw, FP8 MLA KV, 327,680 configured tokens, C4, TP2/EP2/DCP2, no speculation	2x RTX PRO 6000 Blackwell Max-Q	Four matched 32-request finalist cells passed; shared mean visible TTFT −67.84%, request throughput 3.17×; unique request throughput −2.39%; cache reuse 75.58%; final diagnostics 12/12 per arm; vision/context gates passed	Explicit human approval and fresh bounded acceptance; exact r9 remains the rollback	Dossier · Promotion follow-up · Historical campaign

2026-09-18 vision acceptance

Date	Capability	Configuration	Measured GPU	Evidence	Decision	Links
2026-09-18	Eight-image comparison and OCR	GLM EXL3 r9 vision8, FP8 KV, no speculation, configured 327K/C4	2x RTX PRO 6000 Blackwell Max-Q	Four ordinary comparisons, one high-resolution request, six direct preflight checks; matching Pi output has no correlated request provenance	User-authorized eight-image limit; r8 one-image rollback; no concurrency soak	Dossier · Finding
2026-09-18	Image/OCR and text/tools	GLM EXL3 r8 vision, FP8 KV, no speculation, configured 327K/C4	2x RTX PRO 6000 Blackwell Max-Q	Functional; 10/10 images; 260K retrieval passed on retry after initial refusal	User-authorized vision enablement; prior text profile retained	Dossier · Finding

This manually maintained catalog indexes retained, decision-relevant runs. Measured hardware names only the device that executed the workload. PRO relationship prevents a co-resident or protected PRO 6000 from being reported as measured.

Every status and decision in this catalog belongs to its dated evidence. Terms such as current, Primary, rollback, and production describe the published decision at that point in history; they do not report active routes, deployment, placement, or availability.

Research-only recipe intake

These records are indexed for decision reachability but are not benchmark runs and name no measured hardware.

Date	Subject	Evidence	Decision	Dossier / finding
2026-08-15	Qwen3.8 27B external recipe refresh	external-prior; current official vLLM/SGLang recipes, exact-hardware Reddit MTP sweep, X discovery, and community runtime reports	Queue matched official-FP8 MTP=3/4/5 A/B; prepare digest-pinned SGLang official-weight spike; no live or promotion change	Qwen3.8 27B · recipe refresh

Cross-run benchmark interpretations

These records synthesize retained measurements without claiming a new run.

Date	Subject	Evidence	Decision	Dossier / finding
2026-09-03	GLM-5.3-Flash scheduler concurrency versus KV capacity	derivative capacity comparison of the retained 2026-08-29, 2026-08-31, and 2026-09-02 local artifacts; no live request or configuration change	C16 is a short-request scheduler ceiling, not sixteen full windows; keep current SGLang at qualified C1, recognize the corrected rollback's measured C2 long-context headroom, and benchmark any higher current-profile admission before changing the published contract	GLM-5.3-Flash · concurrency interpretation

Failed startup attempts

Date	Configuration	Hardware used for startup	Outcome / boundary	Evidence
2026-09-17	GLM mixed 3.5-bpw EXL3, derived v84, TP2/DCP1, NVFP4 KV, configured 327K/C1	2x RTX PRO 6000 Max-Q, native Linux, 96 GB installed host RAM	Host RAM/swap exhausted before readiness; no inference metrics; rejected configuration, no promotion; baseline restored	GLM dossier · Finding

RTX PRO 6000 runs

Date	Capability / configuration	Measured hardware	Evidence	Decision	Dossier / finding
2026-09-19	Huihui Qwen3.8 compatible-runtime baked MTP3 32K	1x RTX 5090, Windows/WSL2 direct lane	baked quality gates passed; vision 12/12; capacity 12/12 at 181.46 mean decode tok/s and nominal-31K 6/6 at 174.55	qualified bounded profile; clean recreation and router preflight 6/6 passed; no actual deployment assignment disclosed	Qwen3.8 27B · promotion addendum
2026-09-19	Huihui Qwen3.8 compatible-runtime historical comparison set	1x RTX 5090, Windows/WSL2 direct lane	four prior follow-up capacity runs plus two baked capacity runs published to historical evidence	descriptive historical evidence; failed 128-word warmups retained outside performance rows	Qwen3.8 27B · promotion addendum · promotion evidence
2026-09-17	GLM mixed 3.5-bpw EXL3 R7 loader repair, TP2/DCP1, NVFP4 KV, configured 327K/C4	2x RTX PRO 6000 Max-Q, native Linux	C1/C4 functional and repeated bounded quality pass; 216,307 actual-token needle; C4 capacity 4/4 then 3/4, baseline 2/4; strict failures retained	Runnable experimental candidate; no overall win; baseline restored, no-promotion	GLM dossier · Finding
2026-09-14	GLM-5.3-Flash EXL3 4-bpw r7, no speculation; 327,680 total / 65,536 output, C4	2x RTX PRO 6000 Blackwell Max-Q	quality, functional, capacity; MMLU-Pro 90/100 one pass, agentic 30/30, SWE 4/5, strict120 120/120, post-promotion context 9/9	selected September 14 text-only lane; Pi/Hermes/OpenClaw tool checks pass; no exhaustive intelligence winner or full-window concurrency claim	GLM dossier · Finding
2026-09-11	GLM v0.4.3, fixed ormandj weights, 524K/C4	2x RTX PRO 6000 Blackwell Max-Q	functional, capacity, bounded quality; Pi 14/14, retrieval 12/12	user-authorized promotion; 4K output retained	GLM dossier · Finding
2026-09-11	GLM rc14 Pi completion budget 4096 vs 16384; same 393K/C1 backend	2x RTX PRO 6000 Blackwell Max-Q	bounded functional; 16 valid / 24 tasks, eight coding runs invalid	no-promotion; retain 4096	GLM dossier · Finding

Date	Capability	Exact model/configuration	Measured hardware	Evidence	Decision	Dossier / finding
2026-09-26	Single-GPU Swift/Qwen3.8 qualification	Swift Flash IQ3 r3, Swift27 NVFP4 r2, and ThinkingCap AWQ r1; 327,680 total-token configurations, C1 deterministic suites	1x RTX PRO 6000 Blackwell Max-Q per isolated candidate	functional, bounded quality; all final candidates agentic 18/18, image 12/12 plus eight-image 4/4, context 12/12; Flash SWE 3/5, Swift27 2/5 (4 graded), ThinkingCap partial with upper bound 3/5	retain exact dual-GPU GLM r11; candidate configurations rejected/no-promotion; no matched performance claim	Efficient variants · ThinkingCap · GLM · finding
2026-09-19	VoiceChat feasibility stop	mlx-community/NemotronLabs-VoiceChat-11B-4bit@dffd203f; MLX runtime 77a6cfca; no candidate load	Apple M4 Max, 48 GiB unified memory	read-only feasibility; disk arithmetic only, memory unresolved; protected audio diagnostics excluded from candidate latency/quality	whole tool-capable replacement blocked by missing tool-result ingress; no-promotion	Voice LLM MLX · finding
2026-09-09	ormandj v0.4.2 runtime upgrade qualification	GLM-5.3-Flash c3cbb989, candidate image 877ae294, TP2/393,216/C1, P2P enabled	2× RTX PRO 6000 Blackwell Max-Q	functional/thinking/multimodal/high-context pass; 4K n12 and 120K n3 matched cells; strict turnover 58/60 and repeat 57/60	user-selected retain-baseline/no-promotion; exact baseline restored	GLM-5.3-Flash · finding
2026-09-09	Native NCCL P2P transport A/B	GLM-5.3-Flash ormandj W4A16/NVFP4 c3cbb989, exact rc14 image 0c063795, TP2/393,216/C1; only P2P disable 1 to 0 changed	2× RTX PRO 6000 Blackwell Max-Q	direct/routed gates and 380K correctness pass; 4K n12 and 120K n3 matched medians; 380K strict capacity and diagnostic marker failure retained	user-authorized native P2P default; bounded transport decision	GLM-5.3-Flash · finding
2026-09-08	Native Linux / retained WSL comparison	GLM-5.3-Flash ormandj W4A16/NVFP4 c3cbb989, exact rc14 image 0c063795, adaptive EAGLE, FP8 KV, TP2/393,216/C1; native adds disabled NCCL P2P	2× RTX PRO 6000 Blackwell Max-Q	functional, capacity, bounded quality; historical-style n3 decode +20.5–33.0%, endurance 60/60; extended context 128/150 through 376,484 actual tokens (9 empty, 13 incorrect); strict output 0/3 and unique canaries 1/10 excluded	no-promotion; whole-stack migration comparison, no strict finalist qualification; routed context/reasoning limits retained	GLM-5.3-Flash · finding
2026-09-04	Qwen3.8 27B comprehensive SGLang/vLLM optimization and topology campaign	Inferact 6128240e, RadixArk 319f741c, kelnei 29099dc7, incoai DFlash2 dedf8df6; SGLang digest 616a3e97 / source 5f55db35 and vLLM 0.27.1 digest c2f3b1b9; 262,144 context, FP8 KV, K/chunk/compile/Mamba/target/runtime A/Bs	2x RTX PRO 6000 Max-Q under Docker Desktop/WSL2; single TP1, one TP2 service, and two independent TP1 replicas measured	functional, capacity, matched performance, correctness rejection, exact restoration; sustained-output N100: TP1 764.3, TP2 587.9, DP2 1,401.8–1,423.4 aggregate tok/s; all 100/100 canaries; full mean/p50/p95/p99 TTFT, prefill, decode, TPOT/ITL, E2E; RadixArk 746.7 TTFT tradeoff; kelnei MTP2 503.4 versus 315.2 no-spec; 82K/C8 non-interactive	DP2 bounded aggregate-throughput winner; TP2 rejected after 23.1% regression and repeated strict-JSON corruption; RadixArk and kelnei retained tradeoffs; no-promotion; exact GLM TP2 service and authenticated route restored	Qwen3.8 27B · possibility campaign
2026-09-02	GLM-5.3-Flash SWE-bench Verified smoke	ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO@c3cbb989, SGLang 4c2c169b, rc.14 image digest 0c063795, TP=2, FP8 KV, adaptive EAGLE [3,5], 393,216/C1, 4,096 output; thinking enabled; isolated macOS arm64 worker	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	bounded quality and benchmark-infrastructure qualification; fixed django__django-11099 attempted 1/1, officially graded 1/1, resolved 1/1; 11 routed requests; agent 34.216 s and grader 29.076 s; normalized token totals unavailable	adds one-instance repository-agent evidence without changing benchmark guidance; not a full-suite score	GLM-5.3-Flash · SWE smoke
2026-09-02	GLM-5.3-Flash SGLang SM120 393K promotion and client acceptance	ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO@c3cbb989, rc.14 image digest 0c063795, TP=2, FP8 KV, adaptive EAGLE [3,5], 393,216/C1, 4,096 output, image/OCR	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	prior direct functional/capacity/quality/media/endurance qualification; model-only reserve waiver; managed direct and routed promotion gates; real Pi/OpenClaw/Hermes acceptance; 2,543 MiB/card after client workload; exact 524K rollback retained	human-approved published current text/tools/image/OCR profile; Operator Node Pi unchanged pending host-local router credential	GLM-5.3-Flash · promotion
2026-09-02	GLM-5.3-Flash SGLang SM120 W4A16, adaptive-MTP A/B, context-envelope qualification, image/OCR, and endurance	ormandj/GLM-5.3-Flash-W4A16-NVFP4-K32-Experts-FP8-WO@c3cbb989, rc.14 image digest 0c063795, SGLang 4c2c169b, TP=2, FP8 KV, adaptive EAGLE [3,5]; selected 245,760/max-running1, with matched 131K no-spec and policy-infeasible 393K/499,712 controls	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, matched performance, bounded quality, multimodal, endurance, reserve, and restoration; tools 20/20, coding 15/15, image/OCR 12/12, endurance 60/60; selected profile 108.57/93.35/95.00 tok/s decode at nominal 4K/120K/230K and 3,487 MiB post-workload free per card	245,760-token/C1 profile locally verified challenger, no-promotion; 393,216 and 499,712 profiles policy-infeasible; incumbent exactly restored	GLM-5.3-Flash · SGLang qualification
2026-08-31	GLM-5.3-Flash xgrammar fix-forward, matched speculation A/B, 524K context, C2, image/OCR, rollback, and real clients	wrldsuksgo2mars/GLM-5.3-Flash-EXL3-K3-v1@319d66a8, incoai/GLM-5.3-Flash-DFlash2@dc77ff1c, corrected runtime digest 4909e318, TP=2/DCP=2, EXL3 K3 target, FP8 DS-MLA KV, DFlash2 fixed K5, 524,288/maxseq16/batch2,048, router c16, up to 16 images	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, matched performance, bounded quality, rollback, routed/client acceptance; both arms 28/28 functional; 83.08 tok/s at 4K and pooled 69.99 at 240K; C2 nominal 250K 2/2; 2,493,817 KV tokens; real OpenClaw/Hermes/Pi pass	corrected 524K DFlash2 K5 selected as current; former 1M profile first rollback; text/image/OCR, no video; DFlash2 noncommercial boundary	GLM-5.3-Flash · xgrammar qualification
2026-08-30	GLM-5.3-Flash text/tools/image/OCR, 1M context, and K3/K5/chunk optimization	wrldsuksgo2mars/GLM-5.3-Flash-EXL3-K3-v1@319d66a8, incoai/GLM-5.3-Flash-DFlash2@dc77ff1c, digest-pinned Purtell runtime 001a45bd, TP=2/DCP=2, EXL3 K3 target, FP8 DS-MLA KV, DFlash2 K5, 1,048,576/maxseq16/batch2,048, c16, up to 16 images	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, matched performance, bounded quality, routed/client acceptance; tools 20/20, image/OCR 12/12, quality 12/12, exact 950K-target retrieval, 82.1/67.4/67.9 tok/s at 4K/131K/240K; real Hermes/Pi/OpenClaw pass	historical verified; first same-model rollback after the 2026-08-31 fix-forward; K3 alternate and batch4,096 rejection retained; text/image/OCR, no video; DFlash2 noncommercial boundary	GLM-5.3-Flash · optimization and promotion
2026-08-29	GLM-5.3-Flash text/tools/image/OCR, speculation, and 524K context qualification	brandonmusic/GLM-5.3-Flash-tr3-4bpw@5ab363a8, digest-pinned Purtell EXL3/B12x runtime da5cec95, TP=2/DCP=2, NVFP4 DS-MLA KV; vision/text fixed K5, adaptive K1-K5+ReplaySSM, and no-spec controls at 262,144/524,288	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, matched performance, bounded quality; vision/OCR, tools 20/20, high-reasoning coding 15/15, 250K-target / 206,296-actual retrieval, and 72.8/55.7 tok/s at 4K/128K; exact 495,045-token retrieval and 497,976-token tool use on 524K text profiles; adaptive tools 12/20	vision fixed K5 interactive and no-spec 524K maximum-context challenger profiles, no-promotion; adaptive MTP rejected; 0xSero 3.0-bpw watch-only; direct vision candidate retained for hands-on	GLM-5.3-Flash · qualification
2026-08-26	Primary image/OCR/video expansion and comprehensive context sweep	RadixArk Qwen3.8 Flash Next NVFP4 7b719225, digest-pinned SGLang 59f06adc, engine d91c3682, hash-gated PR #36556 QSA, TP=2, 262,144 context, c1, BF16 KV, NEXTN 3/1/4, four images or one video	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, multimodal repeatability, routed/client acceptance; direct 30/30, isolated 27/30 then 30/30, live 29/30 then 28/30; edges 8/8; 25/25 context requests through 245,000 actual prompt tokens; 155.9/114.7/112.9 median decode tok/s at 4K/128K/254K targets; 516,032 server tokens	human-authorized current text/image/OCR/video Primary; 262K/c1, four-image/one-video, thinking-disabled contract	Qwen3.8 Flash Next · vision promotion
2026-08-26	Text Primary QSA-fast/MTP3 fix-forward promotion	RadixArk Qwen3.8 Flash Next NVFP4 7b719225, digest-pinned SGLang 59f06adc, engine d91c3682, hash-gated PR #36556 QSA fast path, TP=2, 262,144 context, c1, BF16 KV, NEXTN 3/1/4	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, matched performance, bounded quality, routed identity and fresh real-client acceptance; tools 20/20 and quality 12/12; MTP3 154.9 tok/s at 4K, 134.1 at 128K, 102.0 at 253,703 actual prompt tokens with 8,192 output request; 2.33x/1.93x matched no-spec decode	human-authorized current text Primary; 262K/c1 thinking-disabled contract; multimodal unpromoted	Qwen3.8 Flash Next · QSA-fast promotion
2026-08-26	Initial text Primary qualification and promotion	RadixArk Qwen3.8 Flash Next NVFP4 7b719225, digest-pinned SGLang 59f06adc, engine d91c3682, TP=2, 262,144 context, c1, no speculation, portable QSA sparse decode plus NCCL logits fallback	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, bounded quality, routed and real-client acceptance; 253,325-token direct/routed retrieval with 8,192 reserve, intelligence 6/6, session 3/3, tools 20/20 plus repeated tools 3/3, Responses, OpenClaw/Hermes/Pi pass; 4K/c1 214.612 ms TTFT and 12.801 tok/s decode	superseded same day by the QSA-fast/MTP3 fix-forward promotion; retained correctness baseline	Qwen3.8 Flash Next · initial promotion
2026-08-21	DeepSeek 0731 Infernal Invocation r18 1M promotion	Official 9e165c30, digest-pinned r18 414ec7d0, B12X W4A8/FP8 compressed MLA KV, TP=2/DCP=1, 1,048,576 context, maxseq8, batch4,096, fixed probabilistic DSpark K5; matched no-spec control	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, matched performance, bounded quality, routed and real-client acceptance; 1,040,063-token retrieval, tools 160/160 plus functional batches, agentic 12/12, c8 short and c2 at 490,861 prompt tokens/request; K5/no-spec decode 142.1/76.4 tok/s at 4K and 129.5/76.3 at 32K; Hermes/Pi/OpenClaw pass	former human-approved text Primary as of 2026-08-26; retained evidence	DeepSeek 0731 · promotion
2026-08-16	DeepSeek 0731 Infernal Invocation r15 393K promotion	Official 9e165c30, digest-pinned r15 f1b13c86, B12X W4A8/FP8 compressed MLA KV, TP=2/DCP=1, 393,216 context, maxseq8, batch4,096, fixed probabilistic DSpark K5; matched target-only control	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2 under WSL2	functional, capacity, matched performance, bounded quality, routed acceptance; K5/no-spec 4K decode 150.0/76.4 tok/s, direct 351,118-token and routed 340,119-token retrieval, repeated quality 12/12, routed tools and OpenClaw-compatible wire paths pass	former human-approved text Primary; r33 393K fixed-port managed rollback at that date; actual Mini OpenClaw turn remained unproven	DeepSeek 0731 · promotion
2026-08-16	Qwen3.8 27B SGLang video and router admission	Official FP8 017b9c7a, digest-pinned SGLang c4271c3, TP=1, 393,216 context, FP8 E4M3 KV, EAGLE 3/1/4, c1, CPU feature transport	1x RTX PRO 6000 Max-Q active; second equal card dormant	functional, deterministic multimodal, routed acceptance; direct 30/30 with video 14/14, live admitted 28/28, overflow/malformed/SSE/tool and Primary regression pass	expand current in place with vision.video, two images/one video fail-closed admission; no model restart	Qwen3.8 27B · video-router finding
2026-08-15	Remote agentic and SWE-bench Verified scout	Official FP8 017b9c7a, digest-pinned SGLang c4271c3, TP=1, 393,216 context, FP8 E4M3 KV, EAGLE 3/1/4, c1; router-only Apple evaluation worker	1x RTX PRO 6000 Max-Q active; second equal card idle	bounded quality; agentic smoke 2/2, scout 16/18, fixed SWE Verified sample 5/5 official-grader resolved	retain current; debug-loop follow-up and larger stratified SWE sample recommended; no route or promotion change	Qwen3.8 27B · agentic/SWE scout
2026-08-15	Qwen3.8 27B SGLang official-FP8 single-service promotion	Official FP8 017b9c7a, digest-pinned SGLang c4271c3, TP=1, 393,216 context, FP8 E4M3 KV, max-running1, 2K chunks, FlashInfer, EAGLE 3/1/4, five GDN states, CPU feature transport	1x RTX PRO 6000 Max-Q active; second equal card empty	functional, capacity, deterministic multimodal, routed and real-client acceptance; 108K retrieval, tools 20/20, direct and routed media 18/18, Responses, Hermes and OpenClaw pass	human-approved current; Primary/general-vision/OCR consolidated on one service; two-image/no-video admission	Qwen3.8 27B · promotion
2026-08-15	Qwen3.8 27B SGLang single-service consolidation A/B	Official BF16 1d4bf0f, official FP8 017b9c7a, and Inferact NVFP4 6128240e; digest-pinned SGLang c4271c3, TP=1, 393,216 context, FP8 KV, max-running1, 2K chunks, FlashInfer, EAGLE 3/1/4, CPU feature transport	2x RTX PRO 6000 Max-Q; BF16 fixed on one card, quantized candidates sequential on the other	functional, capacity, deterministic multimodal; all three pass 18/18 across six image tasks including a two-image comparison; BF16/FP8/NVFP4 media p50 0.915/0.588/0.448 s and 4K decode 62.7/111.4/97.7 tok/s	Official FP8 preferred single-service challenger; NVFP4 lowest-latency third-party alternative; no-promotion; exact current vLLM split restored and routed	Qwen3.8 27B · consolidation A/B
2026-08-15	Qwen3.8 27B SGLang MTP=3 and CPU-transport multimodal A/B	Official FP8 017b9c7a and Inferact NVFP4 6128240e; digest-pinned SGLang c4271c3, TP=1, 393,216 context, FP8 KV, max-running1, 2K chunks, FlashInfer, EAGLE 3/1/4, five GDN states; multimodal arms force CPU feature transport	2x RTX PRO 6000 Max-Q, one independent candidate per equal card followed by a swap	functional, capacity, bounded deterministic quality, bounded multimodal; official FP8 MTP averages 111.3 decode tok/s and NVFP4 98.1; both pass 389K retrieval, intelligence 6/6, session 3/3, tools 3/3, image, and OCR	SGLang MTP/CPU-multimodal challenger, no-promotion; official FP8 wins decode, NVFP4 wins TTFT/prefill; exact current vLLM split restored and readmitted	Qwen3.8 27B · MTP/multimodal qualification
2026-08-15	Qwen3.8 27B SGLang official-FP8/NVFP4 A/B	Official FP8 017b9c7a versus Inferact ModelOpt NVFP4 6128240e; digest-pinned SGLang c4271c3, TP=1, 393,216 context, FP8 KV, max-running1, 2K chunks, FlashInfer, no speculation; models swapped across cards	2x RTX PRO 6000 Max-Q, one independent candidate per equal card	functional, capacity, bounded deterministic quality; both placements pass full gate, tools 20/20, repeated intelligence/session/tools, and 388,979-token retrieval; NVFP4 averages 0.429 s TTFT, 8,409 prefill tok/s, and 57.9 decode tok/s versus 0.554 s, 6,512, and 48.0	NVFP4 text challenger, both no-promotion; SGLang multimodal unqualified on WSL2; exact current vLLM split restored and readmitted	Qwen3.8 27B · SGLang/NVFP4 qualification
2026-08-15	Qwen3.8 27B official-FP8 MTP-depth A/B	Official FP8 017b9c7a, vLLM 3a091411, TP=1, 393,216 context, FP8 KV, maxseq1, batch 4,096, MTP=4 versus MTP=5; settings swapped across cards	2x RTX PRO 6000 Max-Q, one independent candidate per equal card	functional, capacity, bounded deterministic quality; both pass tools 20/20, repeated intelligence/session/tools, and 388,979-token retrieval; same-card MTP=5 decode only 0.4-1.3% above MTP=4 with no E2E win	MTP=4/5 no-promotion; retain current MTP=3; exact FP8/BF16 split restored and readmitted	Qwen3.8 27B · MTP-depth qualification
2026-08-14	Qwen3.8 27B split promotion	Official FP8 017b9c7a text plus official BF16 1d4bf0f2 multimodal, vLLM 3a091411, TP=1 per card, 393,216 context, FP8 KV, maxseq1, batch 4,096, MTP=3; BF16 32 images/request	2x RTX PRO 6000 Max-Q, one independent serve per equal card	functional, capacity, multimodal, client acceptance; FP8 93.6 tok/s at 4K, routed tools 20/20; BF16 routed media 30/30 and 32-image request 1/1; Hermes text/image and OpenClaw Primary/vision passed without fallback	human-approved current split; FP8 text Primary, BF16 explicit general-vision/OCR; TP=2 long-context profiles remain experimental	Qwen3.8 27B · promotion
2026-08-14	Qwen3.8 27B TP/MTP/long-context matrix	Official BF16 1d4bf0f2 and official FP8 017b9c7a, vLLM 3a091411; split TP=1 at 393K and exclusive TP=2 at 393K/600K/1.01M; matched no-MTP/MTP=3, FP8 KV, maxseq1, batch 4,096	2x RTX PRO 6000 Max-Q over PCIe without P2P or NVLink	functional, capacity; all 16 arms passed complete functional gates and cold retrieval at 388,979/598,729/985,107 actual prompt tokens; 4K FP8 MTP 85.9-93.6 tok/s; TP=2 control cut 393K TTFT 35-38%	challenger, no-promotion; TP=2 for prefill/capacity, MTP for decode, 600K/1M batch-like; exact split baselines restored	Qwen3.8 27B · matrix
2026-08-14	Qwen3.8 27B official FP8 1M-context continuation	Official FP8 017b9c7a, vLLM 3a091411, 1,010,000 configured context, TP=1, FP8 KV, maxseq1, batch 4,096, text-only, no MTP/prefix cache	1x RTX PRO 6000 Max-Q; second equal card continued the BF16 serve	functional, capacity; retrieval passed at 316,849, 422,449, and 633,649 actual prompt tokens, then 825,049 at 3/3 with 956.739 s mean E2E; post-stress and restored-control gates passed	stable offline/batch challenger, no-promotion; original 262K FP8 lane restored, no route change	Qwen3.8 27B · 1M continuation
2026-08-14	Qwen3.8 27B official qualification and setting A/Bs	Official BF16 1d4bf0f2 multimodal and official FP8 017b9c7a text, vLLM 3a091411, 262,144 context, TP=1 per split lane; FP8 MTP=3, prefix-cache, and unquantized-KV one-variable arms	2x RTX PRO 6000 Max-Q, one independent card per co-resident serve	functional, capacity, bounded quality, multimodal; both baselines through 241,250 actual prompt tokens, FP8 c5 30/30, BF16 media 30/30; MTP c1 94.8 tok/s, warm 30K-prefix TTFT 0.41 s	challenger, no-promotion; direct baselines healthy, no route change; durable worker agentic/SWE open	Qwen3.8 27B · qualification
2026-08-11	DeepSeek 0731 r33 393K Primary promotion	Official revision 9e165c30, digest-pinned r33 B12X mixed NVFP4-MoE/FP8 path, FP8 DS-MLA KV, DSpark K5, TP=2, 393,216 context, maxseq16, batch 4,096, GPU-only	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2	functional, capacity, bounded quality; direct 359,900-token pass, routed functional pass except legacy long-needle calibration and trivial-prompt reasoning-evidence checks; OpenClaw and Hermes 393K/32K/high client-path passes	human-approved current; routed/OpenClaw/Hermes >300K and SWE score remain open	DeepSeek 0731 · promotion
2026-08-10	DeepSeek 0731 r33 batch-token A/B	Official revision 9e165c30, digest-pinned r33 B12X mixed NVFP4-MoE/FP8 path, FP8 DS-MLA KV, target-only/no-spec, TP=2, 131,072 context, c1; only functional change 8,192 to 4,096 batched tokens	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2	functional, capacity, bounded performance; fresh bracketed starts, 6/6 preflight, matched 119,503-prompt-token pass; activation 1.73 to 1.14 GiB, minimum-rank KV 15.27 to 15.99 GiB; reported KV tokens 283,917 to 553,243 with accounting caveat	4,096 candidate healthy and direct-only at campaign close; next arm GPU-only 393K; no-promotion	DeepSeek 0731 · batch-token A/B
2026-08-10	DeepSeek 0731 r33 quality-first control	Official revision 9e165c30, digest-pinned r33 B12X W4A8 mixed NVFP4-MoE/FP8 runtime, FP8 DS-MLA KV, target-only/no-spec, TP=2, 131,072 context, c1; 393K FP8-KV plus 16 GiB host-offload recipe translated but not loaded	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive TP=2	functional, capacity, bounded quality; preflight 6/6, repeated intelligence/session/tools pass, 119,503 actual prompt tokens at 17.445 s TTFT and 73.86 decode tok/s; context-target calibration caveat retained	priority challenger, no-promotion; no route change	DeepSeek 0731 · r33 control
2026-08-07	DeepSeek 0731 Vision (NVFP4) first load	webbrain-one/DeepSeek-V4-Flash-0731-Vision-NVFP4 rev 3a8f168c, digest-pinned SGLang v0.5.16 image, marlin/marlin kernels, TP=2, KV fp8_e4m3, 4,096 context, --mem-fraction-static 0.97	2x RTX PRO 6000 Max-Q over PCIe without NVLink; exclusive dual-gpu-exclusive	compatibility-only, bounded functional (text lane), bounded negative quality (vision lane)	no-promotion; failed back to the 650K Primary in the same session	DeepSeek 0731 · vision first-load
2026-08-03	Remote context, agentic recovery, and SWE-bench Verified smoke	Official DeepSeek 0731 revision 9e165c30, pinned r16 B12X W4A8/FP8 image, FP8 MLA KV, TP=2, DSpark K5, 650K/maxseq16; Apple evaluation worker	2x RTX PRO 6000 Max-Q over PCIe without NVLink	functional, bounded quality; 8K context 1/1, tool-error retry protocol pass with final-answer fail, SWE-bench Verified 1/1 official-grader smoke	benchmark substrate qualified for scout; no promotion or route decision	DeepSeek 0731 · smoke finding
2026-08-02	DeepSeek 0731 Primary promotion	Official revision 9e165c30, pinned r16 B12X W4A8/FP8 image, FP8 MLA KV, TP=2, DSpark K5, 650K/maxseq16/batch4096; router output cap 32768	2x RTX PRO 6000 Max-Q over PCIe without NVLink; display on AMD iGPU	functional, capacity; Dark Pi, Mini Pi, Mini OpenClaw, oversized-output clamp, exact readiness and exclusive ownership	current, human-approved; 1M removed from client-facing Primary after two fatal client-shaped workspace failures	DeepSeek 0731 · promotion
2026-08-02	DeepSeek 0731 GPU-only 650K/1M Pi qualification	Official revision 9e165c30, pinned r16 B12X W4A8/FP8 image, FP8 MLA KV, TP=2, DSpark K5; 650K/maxseq16/batch4096 and 1M/maxseq1/4/16/batch2048	2x RTX PRO 6000 Max-Q over PCIe without NVLink; display on AMD iGPU	functional, capacity; 640K and 985K retrieval, matched 32K c1 speed, three-tool burst, retained maxseq1 fatal workspace failure	650K/maxseq16 preferred everyday Pi experiment; 1M/maxseq16 preferred deep-session experiment; maxseq4 passing alternative; maxseq1 rejected; all no-promotion	DeepSeek 0731 · 650K/1M Pi qualification
2026-08-02	DeepSeek 0731 native KV offload and 256K	Official revision 9e165c30, derived r16 B12X mmap-unpinned image 331b7925, FP8 MLA KV, TP=2, DSpark K5, 8/16 GiB CPU offload, 262,144 served tokens, c1	2x RTX PRO 6000 Max-Q over PCIe without NVLink	functional, capacity; 128K store/replay, cold 125K/192K/250K ladder, 16 GiB 113,408-token CPU reload, live lifecycle regression	priority challenger, no-promotion; 128K remains preferred performance lane	DeepSeek 0731 · native-offload/256K qualification
2026-08-01	DeepSeek 0731 DSpark and 128K	Official revision 9e165c30, pinned r16 B12X W4A8/FP8 image, FP8 KV, TP=2, DSpark K5 and same-image no-spec control, c1	2x RTX PRO 6000 Max-Q over PCIe without NVLink	functional, capacity, bounded quality; low/high/max preflight, 128K, 27/27 coding-agent attempts, paired speculative A/B, per-card telemetry	priority challenger, no-promotion; DSpark preferred for experiments, 3 GiB reserve failed	DeepSeek 0731 · r16 qualification
2026-08-01	Exclusive TP=2 LLM campaign	DeepSeek V4 Flash 0731, Inkling Small NVFP4, Qwen3.5 122B NVFP4, Nemotron 3 Super 120B NVFP4, and Laguna S 2.1 NVFP4; exact pinned revisions, c1, no speculative decode	2× RTX PRO 6000 over PCIe without NVLink	external-prior, functional, capacity, quality; failed/calibration lanes retained	all no-promotion; DeepSeek is the priority intelligence challenger; production aliases unchanged	Dossiers · dual-PRO campaign · DeepSeek research
2026-07-29	Primary LLM promotion	Agents-A1 official FP8 4d7d5938, FP8 KV, vLLM f25953cc, 262K, c1, thinking off	RTX PRO 6000	functional, capacity, quality	current; Qwen3.5 becomes immediate rollback	Agents-A1 · promotion
2026-07-29	262K multimodal head-to-head	Agents-A1 official FP8 4d7d5938, FP8 KV, vLLM f25953cc; Qwen3.5 122B NVFP4 98915d83, BF16 KV, NGC 26.06; thinking off, c1	RTX PRO 6000	functional, capacity, quality	Agents-A1 wins bounded comparison; Qwen stays current; no-promotion	Agents-A1 · Qwen3.5 · head-to-head
2026-07-28	Multimodal challenger	Agents-A1 BF16 addff08f, official FP8 4d7d5938, and ProtoLabs NVFP4 ff24ba5c; vLLM f25953cc, FP8 KV, 131K; isolated router and FP8 MoE A/B	RTX PRO 6000	functional, capacity, quality	FP8 principal candidate; NVFP4 compact text Pareto; generated tune rejected; no-promotion	Agents-A1 · multimodal qualification
2026-07-28	Primary LLM	Qwen3.5 122B A10B NVFP4, rev 98915d837c4e7c87ac8296d02e89de19b3207e6d, BF16 KV, 262K, c1	RTX PRO 6000	functional, capacity, quality	current	Qwen3.5 · qualification
2026-07-27	Challenger LLM	Agents-A1, rev addff08f1653ee72765c5cf458fe84556bb34f8e, thinking disabled	RTX PRO 6000	functional, capacity, quality	challenger, no-promotion	Agents-A1 · release sweep
2026-07-27	Heavy control	Laguna S 2.1 NVFP4, rev 07614121b31898586430f189d27a25a0be310843	RTX PRO 6000	functional, capacity, quality	rollback	Laguna · release sweep
2026-07-26	Heavy LLM	Laguna S 2.1 NVFP4, pinned nightly, FP8 KV, 262K, thinking disabled	RTX PRO 6000	functional, capacity, quality	rollback	Laguna · qualification
2026-07-18	Heavy LLM	GPT-OSS Puzzle 88B, rev 9c0e0746a0d2218b28cc7b2cb3ce4e1a2f50fdb2, MXFP4/FP8 KV, 131K	RTX PRO 6000	functional, capacity, quality	rollback	Puzzle · enablement
2026-07-17	Heavy LLM	GPT-OSS Puzzle 88B compatibility and qualification sequence	RTX PRO 6000	compatibility-only, functional	no-promotion at that point	Puzzle · qualification
2026-07-17	Heavy LLM	Gemma 4 31B official QAT W4A16, vLLM 0.25.1, FP8 KV	RTX PRO 6000	functional, capacity, quality	rejected for latency	Gemma 4 · optimization
2026-07-16	Heavy LLM	Gemma 4 official/Unsloth 12B, 26B, 31B configurations	RTX PRO 6000	compatibility-only, capacity, quality	no-promotion / rejected	Gemma 4 · vLLM 0.25.1 sweep
2026-07-16	Heavy LLM	Gemma 4 Unsloth NVFP4 follow-up	RTX PRO 6000	capacity, quality	no-promotion	Gemma 4 · follow-up
2026-07-16	Heavy LLM	Gemma 4 chat-template and size bakeoff	RTX PRO 6000	compatibility-only, quality	no-promotion / rejected	Gemma 4 · template bakeoff
2026-07-13	Heavy LLM	Qwen3.6 27B container recipe and first characterization	RTX PRO 6000	functional, capacity	no-promotion	Qwen3.6 · recipe
2026-07-12	Heavy LLM	ThinkingCap Qwen3.6 27B FP8, rev e48255afd77b403446332be0f595868337b36591	RTX PRO 6000	functional, capacity, quality	historical control	Qwen3.6 · promotion-era record
2026-07-12	Heavy LLM	Qwen3.6 27B community NVFP4+MTP, official FP8, Unsloth NVFP4, ThinkingCap	RTX PRO 6000	functional, capacity, quality	no-promotion	Qwen3.6 · variation bakeoff
2026-07-12	Heavy LLM	Qwen3.6 protocol-v2 comparison	RTX PRO 6000	quality	no-promotion	Qwen3.6 · comparison
2026-07-12	Heavy LLM	Qwen3.6 baseline	RTX PRO 6000	functional, quality	no-promotion	Qwen3.6 · baseline
2026-07-12	Heavy LLM	Qwen3.5 122B MXFP4, rev 345839ea666a70f5035672f7c88afcba6281921f	RTX PRO 6000	functional, capacity, quality	no-promotion	Qwen3.5 · MXFP4 benchmark
2026-07-12	Heavy LLM	Nemotron 3 Super 120B and Mistral Small 4	RTX PRO 6000	functional, capacity, quality	no-promotion	Nemotron Super, Mistral · challengers
2026-07-12	Heavy LLM	Nemotron Puzzle 75B recheck	RTX PRO 6000	functional, capacity, quality	no-promotion	Nemotron Puzzle · recheck
2026-07-12	Heavy LLM	GPT-OSS 120B deterministic recheck	RTX PRO 6000	capacity, quality	no-promotion	GPT-OSS 120B · recheck
2026-07-12	Heavy LLM	Laguna XS 2.1 protocol-v2 attempts	RTX PRO 6000	historical-invalid, compatibility-only	rejected	Laguna · eval v2
2026-07-10–11	Heavy LLM	Ornith 35B FP8, MiniMax M2.7 REAP, DeepSeek V4 Flash, Nemotron Puzzle, Qwen3.6 NVFP4+MTP	RTX PRO 6000	functional, capacity, quality, historical-invalid	no-promotion / rejected	Dossiers · Blackwell bakeoff
2026-07-10	Heavy LLM	Qwen3.5 122B NVFP4, NGC 26.04, FP8 KV, 131K	RTX PRO 6000	functional, capacity, quality	no-promotion at that point	Qwen3.5 · candidate record

RTX 5090 runs

Date	Capability	Exact model/configuration	Measured hardware	PRO relationship	Evidence	Decision	Dossier / finding
2026-10-16	Coding/tools functional comparison	Swift15 ff891a1130adfcf46bf745e06f0fa3dcde2e8c82; QUASAR wallawalla47/Qwen3.8-27B-Quasar-NVFP4-NInferV3@fd3d06cea9df58a54d5051aac9545a4a01a922db; matched image 59df48dd99f0d177b73de57ba88eb6457ed7e4b0c5771c45e8976adcd9f5ef67, text-only 32K, configured C4, benchmark requests C1	RTX 5090	unrelated	functional: tools Q 16/18, Swift 18/18; coding Q 119/123 list and 123/123 tuple sensitivity, Swift 120/123; exact restoration verified	no-promotion; no general quality or timing ranking	Qwen3.8 27B · finding
2026-10-05	Swift15 recovered baseline	Swift derivative ff891a11, NInfer bace20dc, image 59df48dd, NVFP4/K8V4/DFlash2 draft7; 262K/C4 configured; 96 GB host/64 GB WSL	RTX 5090	unrelated; separate SWE CPU worker is not GPU measurement	functional, bounded quality/context; 3 preflight repeats, agentic 18/18, context 36/36, images 12/12, official SWE 4/5 in each of two distinct attempts (replay image-pinned); strict capacity 0/12 ineligible; separate service-v2 48/48 at 8K/24K C1/C4	historical pre-change baseline; final cache-bundle promotion recorded in the next row; no fixed-output speed claim	Efficient variants · Finding
2026-10-05	Swift15 host-cache pilot	Same pinned Swift/NInfer image; 2 GiB/4 slots versus 8 GiB/8 slots; device KV 282,112 unchanged, 40 GiB zero-extra-swap bound	RTX 5090, 96 GB host	unrelated	Both 18/18 canaries; 12 matched revisits/arm: retained tokens 124,648→373,954, statistical-median visible TTFT 9.100→5.353 s / E2E 9.227→5.498 s; native nearest-rank p50 separately retained	pilot and fresh regressions passed; expanded profile promoted after final acceptance; no cold-decode/p99 claim	Efficient variants · Finding
2026-10-05	Flash Next Strata initial feasibility	UD-IQ4_XS 38bb39ee, Strata 6f32ec07; 32K/C1 R36/M50 no swap, native IQ MTP4	RTX 5090, 96 GB host/64 GB WSL	unrelated	Five protocol families and images 12/12 pass; service-v2 8K/C1 12/12 canaries, native nearest-rank p50 TTFT 30.406 s/E2E 30.852 s; practical latency gate failed. Instrumented MMQ/fused 3-request diagnostics separate	rejected after bounded diagnostics; larger-context/C4/finalist quality not run; Swift cache bundle promoted after final acceptance	Flash Next · Finding
2026-09-24	ThinkingCap secondary promotion	Exact qualified AWQ revision f8fe157f, vLLM 0.29.0 digest 082ca6f0, MTP3/BF16-KV/UVA3, 131072 total / 10240 output / C1 / two images	1x RTX 5090	unrelated	Initial functional 7/7 and reasoning 2/2; final multimodal retry pass with relay incident retained; C20 pressure 6/20 with fourteen admission timeouts retained	current, human-authorized router promotion; fleet/UI completion tracked separately	Dossier · Promotion
2026-09-24	ThinkingCap Qwen3.8 27B AWQ 128K MTP3 generic-UVA3 candidate	bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ@f8fe157f207a13f977bc3d620ce10a3e9ba5ab11; vLLM 0.29.0 digest 082ca6f0; configured 131072 total tokens, C1, MTP3, BF16 KV, default FP32 SSM, generic UVA3 actual 3.02 GiB, complete vision weights	1x RTX 5090	unrelated	27 functional, 9 thinking, 18/18 vision, limited 30/30 diagnostic; strict unique-cache/canary N100 short-output population: actual inputs 3609–3687, 45 outputs, p50 TTFT 3003.0448 ms, decode 51.885477 tok/s, E2E 3852.9263 ms, aggregate 11.678388 tok/s; text-only scout 2/2 at 32658 and 125838 actual inputs with 4096 cap	User-selected challenger, no-promotion; full 60-case text context passed, all stop, at actual 32658–32689 and 125838–125869 inputs with 4096 reserve; synthetic agentic 2/2, six turns, four tool calls. No route or alias change; exact-limit/one-over, C2, SWE, soak, video, endurance, and mixed-context remain untested. Earlier 32K lane is qualified faster fallback (158.3552 tok/s, 455.4 ms TTFT); whole-recipe result is not a causal MTP/offload comparison	ThinkingCap Qwen3.8 27B · finding
2026-09-24	ThinkingCap Qwen3.8 27B AWQ earlier 32K qualification	bottlecapai/ThinkingCap-Qwen3.8-27B-NVFP4A4-AWQ@f8fe157f207a13f977bc3d620ce10a3e9ba5ab11; vLLM 0.29.0 digest 082ca6f0; 32K/C1 Triton + MTP3 BF16 KV, GPU-resident weights, fraction .87	1x RTX 5090	unrelated	functional, capacity, limited quality; 27 functional, 8 thinking, 18/18 images, 30/30 limited diagnostic, context 60/60, agentic 2/2; strict N100 actual 3609-3687 input / 45 output: 455.4421 ms p50 TTFT, 158.355179 tok/s p50 decode, 61.14585 tok/s aggregate	Qualified faster small-context alternative and exact rollback, no-promotion; no C2, SWE, soak, video, or universal-context claim. Earlier Flash MTP 99/100 is ineligible; FP8 quality 27/30 failed after reasoning-budget exhaustion	ThinkingCap Qwen3.8 27B · earlier finding
2026-09-22	Secondary native vision format diagnostic	qwen38-huihui-ninfer-schemafix-mtp3-160k@18144690, NInfer 70434721, image 22fa89d5; template kwargs omitted under evaluator policy, C1, one image, 1,024-token limit	RTX 5090	unrelated	compatibility-only; 11/12 exact whole-answer assertions; all 12 stop, no reasoning output or transport error	not-qualified, no-promotion; no latency, capacity, cold/warm, or selection claim	Qwen dossier / finding
2026-09-19	Huihui 64K text/tools/vision and native clients	Pinned lyf/Qwen3.8-27B-Huihui-Abliterated-NInfer-NVFP4@18144690, NInfer 70434721, baked digest 44377d56, MTP3/INT8 KV, 65,536 context/C1	RTX 5090	unrelated	functional, quality, descriptive canary-free C1 capacity; direct7/7, routed6/6, vision12/12, Hermes9/9; 24,224 MiB	historical qualified 64K comparison; no endurance or comparative64K speed claim	Qwen dossier / 64K finding
2026-09-12	Efficient fine-tune/pruning screen	Signal27B, Swift27B, Qwopus Flash27B, Minitron20B; pinned Q6_K/Q4_0 KV, llama.cpp a298422d,64K/C1; Signal/Swift MTP3 and no-spec controls	RTX 5090	unrelated	functional, diagnostic quality, failed strict capacity; all tools20/20; strict thinking-on scores9/10,9/10,1/10,3/10 versus incumbent8/10; small and format-sensitive	retain exact262K incumbent; all challengers no-promotion; no qualified full-contract upgrade	Efficient variants · finding
2026-09-15	Managed image/video bringup and Wan graph repair	FLUX.2 high 1024×1024/four-step smoke; Wan v1 bd12b2de and v2 b93170ee, c1	RTX 5090	unrelated	functional image/video artifacts; one image review pass; v1 video independent failure; v2 partial recovery	no-promotion; Wan v2 available=false, quality unverified	FLUX.2 Klein · Wan2.2 · finding
2026-09-03	Qwen3.8 27B quant/speculation bakeoff	Gittensor, cdiamond, QUASAR, CometKim, Red Hat, Telperion, and Unsloth Dynamic V3 pinned checkpoints; SGLang 0.5.18, llama.cpp b10548, NInfer 1455676b, and vLLM 0.27.1; fresh matched no-spec/spec arms at advertised 262K or bounded 64K profiles	RTX 5090	unrelated	functional, capacity, matched performance, bounded quality, compatibility failure, exact restoration; Gittensor target-only 50.9 ms TTFT/244,002 actual prompt; CometKim MTP3 228.0 tok/s but tools 0/3; Unsloth MTP3 137.7 tok/s/tools 20/20 at 64K; cdiamond MTP8 full-context fallback	Gittensor preferred direct TTFT challenger, Unsloth preferred clean 64K speculative arm, no-promotion; dedicated zero-reserve exception only after idle baseline; exact GGUF incumbent restored	Qwen3.8 27B · comparison · finding
2026-09-03	Direct text/tools performance and long context	neroued/Qwen3.8-27B-nvfp4-NInfer@204e3d92, exact .ninfer SHA-256; NInfer e3aeaf8c; no-spec control versus MTP3/lm-head draft; INT8 KV, 252,928 context, c1, thinking disabled	RTX 5090	unrelated	functional, capacity, matched performance, bounded quality, exact restoration; MTP3 median TTFT 0.430 s, decode 165.9 tok/s, E2E 0.720 s versus 0.421 s/75.3 tok/s/1.085 s control; 201,746-token prompt plus 8,192 output cap passed; shared-prefix tools 17/20; 2,354 MiB free	preferred direct text/tools performance challenger, no-promotion; normal 3 GiB reserve, admission, routed/client, broader agentic/SWE, and promotion-grade runtime gates open; exact GGUF incumbent restored	Qwen3.8 27B · qualification
2026-08-28	Production Hermes text-to-image profiles and cold lifecycle	FLUX.2 Klein 4B FP8 5b4408e5, Qwen3 4B encoder/FLUX.2 VAE 5f526678, ComfyUI v0.33.4, graph 991b63b8, fixed four-step draft 512×512 / standard 768×768 / high 1024×1024, c1; Anvil Serving 0.36.0 fix-forward through b46f6ce; real Hermes MCP	RTX 5090	unrelated	functional, routed/client acceptance, bounded capacity and quality, production cutover; 6/8 strict independent reviews pass with two retained draft fidelity/count failures; warm gateway E2E 1.352/1.242/1.650 s; exact same-job cold resume/teardown pass; final server-issued bundle regression 908.936 s E2E and 0.087 s generation; native MCP bytes matched the authenticated resource; arbitrary width and cross-boundary auth controls passed	exact image workflow available=true, promoted=false; three fixed profiles only; no exact count/material guarantee; video remains quality_failed and unavailable	FLUX.2 Klein · production enablement
2026-08-28	Live text-to-image gateway/client acceptance	FLUX.2 Klein 4B FP8 5b4408e5, Qwen3 4B encoder/FLUX.2 VAE 5f526678, ComfyUI v0.33.4, graph 991b63b8, 512×512, four steps, c1; exact Anvil merge 19320de6; cold approval plus real Hermes MCP	RTX 5090	unrelated	functional, routed/client acceptance, bounded quality; two PNGs passed independent prompt-adherence review; artifact full/range/missing/signature/size/hash checks passed; cold/no-job and teardown controls passed	unavailable candidate, no-promotion; broader quality and production cutover open	FLUX.2 Klein · live validation
2026-08-28	Live text-to-video gateway/client acceptance	Wan2.2 TI2V 5B FP16/UMT5 FP8/VAE c4f60d30, ComfyUI v0.33.4 plus VideoHelperSuite 4ee72c06, graph bd12b2de, 512×288, 17 frames, eight steps, 16 fps, c1; exact Anvil merge 19320de6; real Hermes MCP plus A2A replay	RTX 5090	unrelated	functional, routed/client acceptance, negative bounded quality; 117,738-byte H.264 decoded at 17 frames/1.0625 s and passed artifact controls, but independent contact-sheet review failed spatial/prompt adherence	unavailable candidate, no-promotion; exact workflow version blocked on quality	Wan2.2 · live validation
2026-08-28	Text-to-image generation	FLUX.2 Klein 4B FP8 5b4408e5, Qwen3 4B encoder/FLUX.2 VAE 5f526678, ComfyUI v0.33.4/CUDA 13.0/PyTorch 2.13, graph 991b63b8, 512×512, four steps, c1	RTX 5090	unrelated	functional, capacity; decodable PNG, 9.859 s, peak 12,919 MiB from 943 MiB baseline, queue running 1/pending 0	unavailable candidate, no-promotion; perceptual quality and routed/client acceptance open; final worker removed	FLUX.2 Klein · qualification
2026-08-28	Text-to-video generation	Wan2.2 TI2V 5B FP16/UMT5 FP8/VAE c4f60d30, ComfyUI v0.33.4 plus VideoHelperSuite 4ee72c06, graph bd12b2de, 512×288, 17 frames, eight steps, 16 fps, c1	RTX 5090	unrelated	functional, capacity; decodable H.264 MP4, 9.092 s, peak 18,263 MiB from 943 MiB baseline, queue running 1/pending 0	unavailable candidate, no-promotion; perceptual quality, longer clips, and routed/client acceptance open; final worker removed	Wan2.2 · qualification
2026-08-22	Routed real-client acceptance follow-up	Unsloth Qwen3.8 27B GGUF 4ca72078, Q4_0 weights/KV/projector, Q4_0 MTP3, digest-pinned llama.cpp b10548; 262,144 backend context, c1	RTX 5090	unrelated	functional, routed acceptance; OpenClaw and Hermes exact identity/no-fallback plus shell-tool/result continuation pass; wrong-selector fallback negative control detected; route declared stale 131,072-token SGLang/NVFP4 compatibility metadata	short routed tools pass; truthful fingerprint and routed 250K capacity fail; no-promotion	Qwen3.8 27B · 250K GGUF qualification
2026-08-21	250K Fast-tier text/tools/vision qualification	Unsloth Qwen3.8 27B GGUF 4ca72078, Q4_0 weights/KV/projector, no-spec control versus Q4_0 MTP3, digest-pinned llama.cpp b10548; 262,144 context, c1	RTX 5090	unrelated	functional, capacity, matched performance, bounded quality; 253,822 actual prompt tokens plus 8,192 reserve, long-tools at 110,875, tools 20/20, agentic 16/18, 101-turn endurance 3/3, images 18/18; MTP decode 104.1 versus 69.1 tok/s	FAST-TIER challenger, no-promotion; Q6_K+same-MTP mathematically disqualified; exact pre-test 128K baseline restored	Qwen3.8 27B · 250K GGUF qualification
2026-08-21	Recipe research and same-checkpoint MTP3/ReplaySSM A/B	RadixArk Qwen3.8 27B NVFP4 554ebba9, no-speculation baseline versus native MTP 3/1/4 plus linear ReplaySSM on digest-pinned SGLang f825d729; TP=1, FP8 E4M3 KV, c1, 131,072 declared context	RTX 5090	unrelated	functional, capacity, matched performance, current external-source registry; candidate decode +80.5% at 4K and +67.9% at 64K, tools 20/20, but only 70,231 KV tokens and 64K E2E +1.9% slower	MTP3/ReplaySSM rejected as 128K replacement; external EXL3/NInfer/vLLM leads remain research-only; exact baseline restored at 105,649 prompt tokens, no-promotion	Qwen3.8 27B · recipe research
2026-08-21	DFlash2 compatibility and context-capacity gate	RadixArk Qwen3.8 27B NVFP4 554ebba9 plus incoai DFlash2 dedf8df6, digest-pinned SGLang f825d729, TP=1, configured 262,144 context, FP8 E4M3 KV, c1, DFLASH/8 draft tokens; exact float32/extra_buffer_lazy arm plus BF16 single-slot tuning ladder	RTX 5090	unrelated	functional, capacity; exact arm allocated 24,347 KV tokens and rejected 105,649 prompt tokens; tuned BF16/no-radix/no-prefill-graph arm allocated 70,262 and passed coding, JSON, 49,549-token retrieval, tools 20/20	Both DFlash2 arms rejected as 128K replacements, no-promotion; exact stock 128K candidate restored and passed the complete gate	Qwen3.8 27B · DFlash2 qualification
2026-08-20	Chat-template A/B	RadixArk Qwen3.8 27B NVFP4 554ebba9, digest-pinned SGLang c4271c3, TP=1, 131,072 context, FP8 E4M3 KV, c1, no MTP; stock versus Sharp v22.1 3dc34df	RTX 5090	unrelated	functional, bounded diagnostic quality; Sharp preflight pass, thinking-enabled 24/30 for both with Sharp +10.8% completion tokens/+10.7% latency, thinking-disabled behavior stock 18/18 versus Sharp 15/18 with Sharp -5.1% tokens	Sharp v22.1 rejected, no-promotion; exact stock 128K candidate restored	Qwen3.8 27B · template A/B
2026-08-17	Long-context multimodal boundary	RadixArk Qwen3.8 27B NVFP4 554ebba9, digest-pinned SGLang c4271c3, TP=1, 131,072 context, FP8 E4M3 KV, c1, no MTP, CPU feature transport	RTX 5090	unrelated	functional, bounded deterministic quality; retrieval at 119,675 actual prompt tokens, tools 20/20, media 30/30, boundary 4/4 at eight images / two videos	retained direct 128K challenger, no-promotion; 64K recipe retained as rollback	Qwen3.8 27B · 128K qualification
2026-08-17	Computer-use perception / vision / video	RadixArk Qwen3.8 27B NVFP4 554ebba9, digest-pinned SGLang c4271c3, TP=1, 65,536 context, FP8 E4M3 KV, c1, no MTP, CPU feature transport	RTX 5090	unrelated	functional, capacity, bounded deterministic quality; 60K retrieval, tools 20/20, direct image/OCR/video pass, corpus 30/30 with video 14/14 and mixed 4/4	qualified 64K baseline; now retained as 128K rollback, no-promotion	Qwen3.8 27B · 5090 qualification
2026-07-28	STT	Parakeet TDT 0.6B v3 baseline	RTX 5090	protected/co-resident	quality, capacity	current	Parakeet · ASR qualification
2026-07-28	STT	Qwen3-ASR 0.6B, rev 5eb144179a02acc5e5ba31e748d22b0cf3e303b0	RTX 5090	protected/co-resident	quality, capacity	challenger, no-promotion	Qwen3-ASR · ASR qualification
2026-07-28	STT	Nemotron 3.5 ASR, rev f3d333391852ba876df169dcc9ba902d25b6ab0b	RTX 5090	protected/co-resident	quality, capacity	rejected	Nemotron ASR · ASR qualification
2026-07-27	Omni/vision	Nemotron Nano Omni 30B NVFP4, rev dc5f0b0bfddf8b6e0f5891475be9af05b80126fe	RTX 5090	protected/co-resident	functional, capacity	current topology	Nemotron Omni · Omni qualification
2026-07-27	Omni/voice	Qwen2.5-Omni 3B, rev f75b40e3da2003cdd6e1829b1f420ca70797c34e; Parakeet; Kokoro	RTX 5090	protected/co-resident	functional, capacity	challenger, no-promotion	Qwen Omni, Parakeet, Kokoro · co-resident stack
2026-07-16	Fast LLM	Gemma 4 E2B W4A16, E4B, and controls	RTX 5090	topology-only	quality, capacity; E2B preflights through 120K passed, strict timeout triage 0/3	E4B retained; E2B no-promotion	Gemma E2B · Gemma E4B · template bakeoff
2026-07-13	Fast LLM	Gemma 4 E4B Fast	RTX 5090	topology-only	functional	historical current	Gemma E4B · router promotion
2026-07-10–11	Fast/Omni	Nemotron Nano/Omni 30B, Gemma 4 31B failed load, Qwen3.5 35B GGUF, Gemma E4B GGUF	RTX 5090	topology-only	functional, capacity, quality, historical-invalid	no-promotion / rejected	Qwen3.5 35B · Gemma 4 variants · RTX 5090 dossiers · Blackwell bakeoff

Apple Silicon runs

2026-10-06 English voice campaign

All rows use one Apple M4 Max 40-core GPU/48 GiB host and direct component endpoints. Exact revisions and runtime/source bindings are in the finding and recipe. Control loaded checkpoint revisions remain unattested. Percentiles are client endpoint measurements, not physical playback or end-to-end voice.

Exact configuration	Functional / quality evidence	Scope, failure and decision	Dossier / finding
Parakeet Metal control	human C1 WER 3.064%,p95 81.68 ms	24 human×3; critical error; retained control only	Parakeet · STT
Qwen3-ASR 1.7B MLX q8	human C1 WER 2.786%,p95 314.25 ms	same population; paired CI overlaps zero/critical error; no promotion	Qwen · STT
Granite TurboCTC FP16	human C1 WER 3.343%,p95 40.16 ms	same population; point regression/critical error; no promotion	Granite · STT
Kokoro 0.8.2 MPS streaming	60/60; first PCM p50 1026.71 ms	60 distinct warm+separate cold; ASR proxy/no human winner	Kokoro · TTS
Breeze BF16 streaming	59/60	resource abort; partial timings not ranked	Breeze · TTS
Breeze q8 R2 streaming/cache0/750	60/60; first PCM p50 1527.19 ms	full warm component; fidelity/listening/client gates open	Breeze · TTS
Breeze q8 R3 buffered/cache0/750	60/60; first PCM p50 5458.13 ms	435 resource samples/growth 0; no human/client qualification	Breeze · TTS
Fish q8 R2 buffered/cache0/1024	56/60	resource growth 728099718B; four failed; no partial ranking	Fish · TTS
Baseline 4B	semantic 132/144; historical 36/36; tools 0/3	strict 8K0/3 eligible; retain existing control only	MLX · 4B
Stock 9B off	preflight 9/9; semantic 131/144; historical 33/36	strict 8K0/3 eligible; no promotion	MLX · 9B
Stock 27B off	no native inference	startup 6.19 GB swap growth; post-abort probe transport-only	MLX · abort
Private cache0 27B off	preflight 9/9; semantic 138/144; historical 36/36	strict 8K0/3 eligible, substantive arithmetic/private fixture	MLX · off
Private cache0 27B low	preflight 9/9; semantic 144/144; historical 33/36	strict 8K0/3 eligible; prefix canary/capture failure; no promotion	MLX · low
Private cache0 27B medium	preflight 9/9; semantic 144/144; historical 31/36	strict 8K0/3 eligible;2 reasoning-only/1 capture failure; no promotion	MLX · medium

Static compatibility stops remain separate: Granite q8 group-size mismatch and English Nemotron unsupported pinned loader. They have no latency/quality score. Both medium phases bind the same original pre-up baseline/PID with a verified no-inference handover; resource work closure does not assert correctness or qualification.

Date	Capability	Exact model/configuration	Measured hardware	Evidence	Decision	Dossier / finding
2026-09-19	Swift / stock Qwen3.8-27B artifact feasibility screen	Pinned Swift and stock GGUF Q6_K / Q4_K_M candidates plus family-matched F16 projector; conversion runtime llama.cpp b10896 fa676981, managed binary/config unverified	Apple M4 Max 40-core GPU, 48 GiB unified memory	read-only disk-policy screen only; all four uncached pairs shortfall the retained reserve/allowance before temporary files; memory containment unresolved	investigation stop; no download, trial, benchmark, cleanup, route, or promotion; no-promotion	Efficient variants · Qwen3.8 27B · finding
2026-09-08	Local voice LLM refresh and TTS runtime update	Qwen3.5-9B MLX 4-bit 8b2b98c00a6b4d291155e4890773ca8f769aee53; Qwen3.8-27B 3e6447f082e89cc7f0bc6e5441afd38dfce760ff; Qwen3.6-35B-A3B 38740b847e4cb78f352aba30aa41c76e08e6eb46; MLX-LM 0.31.3 / MLX 0.31.2; Kokoro FastAPI 0.8.2 58b08a915b3463cb76e376a2867e04f9d828f4df	Apple M4 Max laptop, 48 GB unified memory	9B functional 6/6; diagnostic quality 9/12 and patchformat 0/3; spoken 33/36 strict; strict 4K/C1 capacity 0/10 with canaries 10/10; Qwen3.8 compatibility-only 5/6 groups, tools 1/3; Qwen3.6 compatibility-only 5/6 strict JSON fail but spoken 36/36 and one accepted SDK session; Kokoro 0.8.2 one-sample WER 0.0/2139.31 ms and final warm Realtime acceptance	no-promotion for LLMs; no performance headline or live LLM change; candidates unloaded; Kokoro 0.8.2 and versioned TTS/proxy definitions deployed with four endpoint checks HTTP 200	Voice LLM MLX · finding

Related views

• RTX PRO 6000 hardware page
• RTX 5090 hardware page
• Apple M4 Max hardware page
• Model dossiers
• RTX PRO 6000 mention audit
• Chronological findings