Skip to content

RTX PRO 6000 benchmark view

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.

2026-09-18 update: GLM EXL3 vision passed 10/10 synthetic image attempts plus routed acceptance and a matching Pi output transcript on this GPU pair. Configured 327K/C4 retained; 260K retrieval passed on retry after an initial refusal. No new speed or broad quality ranking. The prior text-only record below is historical.

2026-09-19 current decision: GLM-5.3-Flash r10 APC received explicit human approval and fresh bounded acceptance on this pair. The 32-request shared-prefix finalist reduced mean visible TTFT 23.70 s → 7.62 s (−67.84%) and raised requests/s 0.146 → 0.462; unique-prefix requests/s changed 0.144 → 0.141 (−2.39%), inside the frozen 5% gate. The historical campaign retains the 75.58% cache-reuse measurement, failures, and restored-r9 close. Exact r9 is the rollback. This is not a full-window concurrency, video, soak, or broad intelligence result.

2026-09-14 update: GLM Flash EXL3 r7 no-spec was the selected September 14 text-only lane: 327,680 total tokens with a 65,536-token output reserve at C4, 90/100 one-pass MMLU-Pro, agentic 30/30, SWE 4/5, strict120 120/120, and 9/9 post-promotion 258K-class context. The one matched no-spec C4 population measured 38.908 tok/s median decode. This is not an exhaustive intelligence ranking, a general speed claim, or proof of concurrent full-window capacity; fresh boot/reboot coverage is absent. Earlier deployment descriptions below are historical.

Hardware: 2× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 96 GB each (192 GB aggregate), sm_120. Host: Primary Node; native Linux measurements on 2026-09-09, with retained Windows 11/Docker Desktop/WSL2 history. Last locally measured: 2026-09-19. Evidence interpretation reviewed: 2026-09-19.

Side-by-side speed and recipe links for every configuration measured on this card or both cards in TP=2: model comparison table.

This page's result tables contain only measurements made on one or both PRO 6000 cards. Tests that merely kept a card running or described its topology belong in the mention audit, not the result tables.

Recorded topology and service history

The two GPU roles are symmetric. Split mode can place independent workloads on the cards. Exclusive TP=2 mode assigns both roles to one declared owner and blocks every other inference workload until the mode is left; the cards are connected over PCIe without NVLink, so 192 GB is aggregate rather than unified memory. The public 2026-09-02 promotion finding records the pinned ormandj GLM-5.3-Flash SGLang W4A16/NVFP4 adaptive-MTP profile as the 2026-09-02 exclusive TP=2 text/image/OCR reference at 393,216 tokens and C1. It passed the complete direct qualification, managed and routed promotion gates, and real Pi/OpenClaw/Hermes acceptance under an explicit model-only reserve waiver. The corrected 524K EXL3 K3 plus DFlash2 K5 profile was its immediate same-model rollback. The 2026-08-26 RadixArk Qwen3.8 Flash Next NVFP4 profile was its immediate video-capable rollback. The retained September 14 rollback is the original Qwen3.8 27B FP8 path; it was not exercised after the 2026-09-14 promotion. The 2026-09-04 Qwen3.8 27B campaign measured current-SGLang TP1, TP2, and two independent TP1 replicas, plus Inferact/RadixArk targets and kelnei/vLLM MTP2/no-spec. DP2 won the bounded aggregate-throughput workload; TP2 was slower than TP1 and failed strict JSON. Exact baseline restoration passed; the published service and route did not change. The earlier Qwen3.8 27B single-service and split profiles and the DeepSeek Infernal Invocation r18/r15 promotions remain retained history. Earlier findings also retain the r16 650K promotion, 262K retune, and r27 image-upgrade history. Active assignments remain private operator state; Companion Node is model-free in the reference topology and reaches Dark remotely.

Native Linux migration comparison (2026-09-08)

The same physical pair, exact GLM-5.3-Flash model/image and TP2/393,216/C1 profile measured warm direct median decode of 149.02/125.03/124.46/120.29 versus retained WSL 112.07/96.17/102.42/99.79 tok/s at nominal 4K/120K/262K/380K. These n3 shared-prefix, variable-short-output cells show 20.5–33.0% higher decode with long effective prefill within approximately ±2.3%. Endurance completed 60/60 at 142.64 versus 102.19 tok/s (+39.6%). Coding 15/15, image corpus 12/12, agentic 30/30 and official SWE smoke 1/1 are bounded quality evidence; SWE agent time increased to 115.69 from 34.22 seconds. The extended direct context suite passed 128/150 through 376,484 actual prompt tokens, with 9 empty length-terminated and 13 incorrect visible answers; nominal 8K/32K/131K/262K/380K buckets passed 27/18/24/30/29 of 30. This non-monotonic default-thinking curve has no matched Windows arm; its threshold-derived 8,192 effective-context field is not a physical cap or Linux-caused regression. Strict output 0/3 and unique canaries 1/10 are excluded failed populations. Routed 380K admission and reasoning evidence remain failures; the direct image phrase-sensitive miss is retained. The native stack adds NCCL_P2P_DISABLE=1 and changes OS/driver/cache history, so this is not an OS-only causal experiment or strict finalist qualification. no-promotion; prior promotion decisions remain dated history. See the finding and artifacts.

Native NCCL P2P transport A/B (2026-09-09)

Latest functional measurement (2026-09-11): GLM Pi completion-budget comparison on this native Max-Q pair. Same rc14/393K/C1 backend; 16 valid runs and eight invalid coding runs. No output-default or hardware-setting change; this is not throughput qualification.

The 2026-09-09 pinned GLM TP=2/C1 native recipe retained P2P enabled after the bounded transport A/B: 4K n12 and 120K n3 matched cells lowered median TTFT by about 9% and E2E by 7.4–8.9%, while decode changed by +1.4–2.1%. Both 380K strict capacity cells failed the required 32-word count and are excluded from performance claims. See the finding and evidence.

September 17 mixed-quant startup failure

A mixed 3.5-bpw GLM EXL3 startup on the dual Max-Q host exhausted host RAM and all 8 GiB swap before readiness. The two workers reached approximately 77 GiB combined resident memory; the desktop session shut down. No candidate inference or GPU-capacity result was produced. The exact 4-bpw baseline was restored and passed direct and routed short checks. This is a host-loader and containment failure, not proof of a model-quality defect; no promotion occurred.

September 17 bounded loader fix-forward

The R7 streaming fix-forward loaded all 120 shards under 60 GiB RAM/no additional swap and passed C1/C4 functional and repeated C4 quality scouts. It retrieved exact ZEBRA at 216,307 actual input tokens. One strict C4 capacity cell passed 4/4 at 98.17 tok/s for itself, but the next failed at 3/4 (65/64 code words); the baseline also had a 2/4 strict failure (63/64). No overall win, full 327K/C4 context, or repository/session/client qualification is established. The baseline remains selected and the candidate is no-promotion.

Recorded promotion, rollback, and challenger state

The 2026-09-09 ormandj v0.4.2 runtime candidate passed direct gates and bounded matched capacity, but strict turnover was 58/60 then 57/60. The user selected retain-baseline/no-promotion; the exact baseline was restored. See the dated finding.

Order Model Decision Contract
1 GLM-5.3-Flash EXL3 r7 no-spec historical, selected 2026-09-14 text-only lane EXL3 4-bpw, 327,680 total / 65,536 output tokens at C4; one-pass MMLU-Pro 90/100, agentic 30/30, SWE 4/5, strict120 120/120, and context 9/9 through 255,647–255,672 prompt tokens plus reserve. Pi, Hermes, and OpenClaw tool checks pass; no full-window concurrency soak or fresh boot/reboot proof.
2 GLM-5.3-Flash SGLang W4A16 historical 2026-09-02 human-approved 393K/C1 profile Exact rc14 SGLang image and W4A16/NVFP4 checkpoint at exclusive TP=2, 393,216 tokens, C1, 4,096 maximum output, FP8 KV, adaptive EAGLE, image/OCR, and explicit thinking control; direct, managed, routed, and real-client gates passed under a recorded model-only reserve waiver
3 GLM-5.3-Flash historical same-model rollback EXL3 K3 target plus DFlash2 K5 at exclusive TP=2, 524,288 tokens, router c16, up to 16 images, and 8,192 maximum output; both matched arms passed 28/28 functional observations, C2 nominal 250K completed 2/2, and exact router plus real Hermes/Pi/OpenClaw acceptance passed; no video; noncommercial draft boundary
4 Qwen3.8 27B FP8 retained September 14 rollback Original FP8 text-only rollback path; declared but unexercised after the 2026-09-14 promotion
5 Qwen3.8 Flash Next historical 2026-08-26 video-capable rollback RadixArk ModelOpt NVFP4 retains text/image/OCR/video evidence at exclusive TP=2, 262,144 tokens, router concurrency one, four-image/one-video admission, and an 8,192-token output reserve; direct vision 30/30, live repeats 57/60 strict, edges 8/8, full-reserve capacity, routed identity, and real-client acceptance passed
6 Qwen3.8 27B historical DP2 throughput winner, TP2 rejected, no-promotion Current SGLang/Inferact NVFP4 plus DFlash2 K12/chunk1K: two TP1 replicas 1,401.8–1,423.4 aggregate tok/s at C16 versus one TP1 764.3 and TP2 587.9; 100/100 canaries, but TP2 strict JSON failed twice; RadixArk is the TTFT tradeoff and kelnei/vLLM MTP2 is an alternate-runtime gain; exact GLM baseline restored
7 DeepSeek V4 Flash 0731 former current, 2026-08-21 r18 promotion Retained Infernal Invocation r18 B12X/DSpark K5 TP=2/1,048,576 promotion evidence
8 DeepSeek V4 Flash 0731 former current, 2026-08-16 r15 promotion Retained Infernal Invocation r15 B12X/DSpark K5 TP=2/393,216 promotion evidence
9 Qwen3.8 27B former single service and split Retained official-FP8 SGLang text/image/OCR/video profile and FP8/BF16 vLLM split recipes
10 Qwen3.5 122B retained qualified recipe Historical qualification evidence
11 Agents-A1 plus Omni historical managed split Agents-A1 retains FP8 text/image/video evidence
12 Laguna S 2.1 retained historical recipe Thinking disabled
13 GPT-OSS Puzzle 88B retained historical recipe Strict unified-diff caveat

Comparable quality and context

Exclusive dual-card TP=2

Candidate Repeated quality Capacity and context evidence Decision
GLM-5.3-Flash EXL3 r7 no-spec, TP=2/C4 MMLU-Pro 90/100 one pass; agentic 30/30; frozen SWE 4/5; Pi/Hermes/OpenClaw tool checks pass 327,680 total / 65,536 output; 9/9 through 255,647–255,672 actual prompt tokens plus reserve; concurrent full-window capacity unmeasured September 14 text-only reference; strict120 120/120; one C4 no-spec median 38.908 tok/s; no full-window soak or fresh boot/reboot proof
GLM-5.3-Flash ormandj W4A16/NVFP4, SGLang adaptive EAGLE, TP=2, 393K/C1 thinking-disabled preflight 10/10, tools 20/20, long needle/tool, and image/OCR; thinking-enabled contract; coding 15/15, media 12/12, endurance 60/60; managed/routed gates and real Pi/OpenClaw/Hermes pass; fixed SWE-bench Verified smoke officially resolved 1/1 through 11 routed requests 304,491 actual prompt tokens at the 380K target; median decode 112.07/96.17/102.42/99.79 tok/s and effective prefill 16,729/5,749/5,579/5,457 tok/s at 4K/120K/262K/380K; 2,101 MiB/card after qualification and 2,543 after client work; SWE agent/grader 34.216/29.076 s historical 2026-09-02 text/image/OCR selection; C1 and 4,096 output; explicit model-only zero-reserve waiver; one-instance SWE smoke only; 240K conservative fallback; 499K rejected/unverified
GLM-5.3-Flash EXL3 K3 target plus corrected DFlash2 K5, TP=2, 524K/c16 intelligence 6/6, session 3/3, tools 20/20 plus repeated tools 3/3; image/OCR; authenticated route and fresh Hermes/Pi/OpenClaw acceptance 206,296-actual-token retrieval; 2,493,817 KV tokens / 4.76 complete windows; median decode 83.08 tok/s at 4K and pooled 69.99 at 240K; C2 nominal 250K 2/2 historical rollback; K5/batch2,048; 16 images, no video; DFlash2 noncommercial boundary
Qwen3.8 Flash Next RadixArk NVFP4, SGLang QSA-fast MTP3, TP=2, 262K/c1 thinking-disabled intelligence 6/6, session 3/3, tools 3/3; direct media 30/30; isolated router 27/30 then 30/30; live router 29/30 then 28/30; edge suite 8/8; exact routed identity and fresh OpenClaw/Hermes/Pi acceptance 253,703 actual prompt tokens with 8,192 output request; 516,032 maximum server tokens; 6.275 GiB KV per rank; six-size sweep 25/25; median decode 155.9/114.7/112.9 tok/s at 4K/128K/254K targets and 102.0 at full reserve historical text/image/OCR/video rollback; hash-gated PR #36556 SM120 fast path; c1, four images or one video
GLM-5.3-Flash TR3/EXL3 4 bpw, fixed K5/no-spec, TP=2 vision/OCR, tools 20/20, agent protocols, and high-reasoning bounded coding 15/15 pass; adaptive arm tools only 12/20 and rejected vision fixed K5 72.8/55.7 tok/s decode at 4K/128K, exact retrieval at a 250K target / 206,296 actual prompt tokens, and 560,866 KV tokens; 524K text arms pass 495,045-token retrieval and 497,976-token tool use; no-spec reports 1,603,111 KV tokens historical GLM starting point and GLM-specific rollback evidence
DeepSeek V4 Flash 0731, Infernal Invocation r18 B12X + DSpark K5, batch 4,096, maxseq8, 1M complete direct and routed functional gates; clean post-reload gate; intelligence/session/tools 12/12; additional structured tools 160/160; real Hermes/Pi/OpenClaw acceptance K5/no-spec 4K decode 142.1/76.4 tok/s and 32K decode 129.5/76.3; calibrated 1,040,063 actual prompt tokens pass; c8 short 8/8 and c2 at 490,861 prompt tokens/request; 1,323,176 KV tokens / 1.26 full windows former human-approved text Primary; retained evidence
DeepSeek V4 Flash 0731, Infernal Invocation r15 B12X + DSpark K5, batch 4,096, maxseq8, 393K full direct functional gate; repeated tools/session/unified-diff/timeout 12/12; authenticated routed tools and OpenClaw-compatible Anthropic basic/tool paths pass K5/no-spec 4K decode 150.0/76.4 tok/s and 32K decode 119.245/76.767; direct 351,118 and routed 340,119 actual prompt tokens pass; c8 short 8/8 and c2 long 2/2; 797,689 KV tokens / 2.03 full windows former human-approved text Primary; retained evidence; actual Mini OpenClaw turn remained open
DeepSeek V4 Flash 0731, r33 B12X + DSpark K5, batch 4,096, maxseq16, 393K prior repeated high-reasoning suite retained; routed functional checks passed except legacy long-needle calibration and trivial-prompt reasoning-evidence checks; OpenClaw/Hermes 393K/32K/high client paths passed direct 238,507/339,310/359,900 actual prompt tokens passed; largest 65.2 s TTFT and 5,599 effective prefill tok/s; engine 725,543 KV tokens / 1.845 full windows historical Primary; now managed TP=2 rollback; routed/OpenClaw/Hermes >300K and SWE score remain open
DeepSeek V4 Flash 0731, r33 B12X target-only/no-spec, batch 4,096 functional preflight 6/6 at high reasoning; same repeated broad-quality suite not rerun 119,503 actual prompt tokens; 17.364 s TTFT, 7,344 effective prefill tok/s, 75.20 decode tok/s; minimum-rank KV 15.99 GiB; engine reports 553,243 KV tokens with unresolved byte/token-accounting caveat healthy direct-only A/B winner for capacity; next arm GPU-only 393K; no-promotion
DeepSeek V4 Flash 0731, r33 B12X target-only/no-spec intelligence 6/6, session 3/3, tools 3/3 at high reasoning; functional preflight 6/6 119,503 actual prompt tokens; 17.445 s TTFT, 7,537 effective prefill tok/s, 73.86 decode tok/s; 283,917-token GPU KV allocation; context-target calibration caveat retained priority challenger, no-promotion; 393K FP8-KV plus 16 GiB host-offload recipe translated but not loaded
Qwen3.5 122B NVFP4 intelligence 6/6, session 3/3, tools 3/3 32K 12/12 at 2.32 s TTFT and 67.5 tok/s decode; 128K 4/4 at 14.59 s and 65.0 tok/s TP=2 no-promotion; single-card profile remains rollback
Nemotron 3 Super 120B NVFP4, TP=2 + EP=2 intelligence 6/6, session 3/3, tools 3/3 32K 12/12 at 2.84 s and 59.5 tok/s; 60K 4/4 at 5.58 s and 60.0 tok/s no-promotion
Laguna S 2.1 NVFP4 intelligence 6/6, session 3/3, tools 3/3 32K 12/12 at 1.97 s and 70.9 tok/s; 240K 4/4 at 31.85 s and 66.0 tok/s TP=2 no-promotion; single-card profile remains rollback
DeepSeek V4 Flash 0731, r16 B12X + DSpark K5 coding/intelligence/session/tools 27/27; low/high/max functional gates pass 128K pass; warmed 125,785-token row 19.44 s TTFO, 23.81 s visible TTFT, 128.9 tok/s decode; matched 4K decode 130.7 tok/s priority intelligence challenger, no-promotion; DSpark preferred for experiments, both lanes fail 3 GiB reserve
DeepSeek V4 Flash 0731, r16 B12X + DSpark K5, GPU-only Pi contexts 650K/maxseq16 passed the low-reasoning gate plus Dark Pi, Mini Pi, and Mini OpenClaw high-reasoning smokes; 1M retained two fatal client-shaped workspace failures 640K retrieval 120.6 s; matched 32K decode 141.6 tok/s at 650K; 1M qualification reached 985K before later client failures 2026-08-02 promotion record; later retuned to 262K and upgraded to r27; 1M experimental only; 3 GiB reserve explicitly waived
DeepSeek V4 Flash 0731, remote AI-MBP25 benchmark worker 8K context 1/1; tool-error retry protocol passed but final answer failed; SWE-bench Verified official-grader smoke resolved 1/1 6,102 observed prompt tokens in the 8K context case; larger buckets not attempted benchmark substrate qualified for scout; no promotion change
DeepSeek V4 Flash 0731, r16 B12X + DSpark K5 + native KV offload full functional preflight passes; 128K and 256K CPU reload proven 8 GiB cold 249,573-token row: 43.75 s TTFO, 45.58 s visible TTFT, 5,705 effective prefill tok/s, 135.2 tok/s decode; 16 GiB exact 113,674-token replay: 113,408 external hits, 1.002 GB CPU-to-GPU, 0.825 s TTFO; managed mmap cleanup passes capacity extension, no-promotion; no 256K per-card reserve sample
DeepSeek V4 Flash 0731, earlier SGLang lane intelligence 6/6, session 3/3, tools 3/3 at low reasoning 32K 11/12; 2.70 s TTFO, 29.11 s first-visible TTFT, 11.5 tok/s combined reasoning/visible decode retained low-reasoning point-in-time lane; one reasoning-only exhaustion; see the r16 and r15 rows for later DSpark evidence
Inkling Small NVFP4 intelligence 6/6, session 3/3, tools 3/3 at low reasoning 32K 12/12; 2.79 s TTFO, 4.63 s first-visible TTFT, 73.5 tok/s combined reasoning/visible decode; reasoning-off lane also 12/12 no-promotion; reasoning-off Responses caveat retained

The historical rows below used both physical cards as measured hardware, exclusive ownership, one admitted request, and no co-resident inference. See the dated campaign for exact revisions, raw artifacts, protocol differences, and failure records.

DeepSeek 0731 research priority and r16 result

DeepSeek identifies 0731 as a re-post-trained official Flash generation with the same 284B/13B-active target architecture as Preview and a bundled DSpark draft module. Artificial Analysis independently scores the max-effort model at 50, number 3 of 101 comparable models, but reports 210 million evaluation output tokens. That verbosity signal is consistent with the earlier 11/12 run's reasoning-only exhaustion and makes reasoning-budget policy a first-class capacity gate.

The follow-up translated a pinned r16 B12X/InstantTensor recipe to WSL2 and qualified the official release revision at 131K with DSpark K5. Low, high, and max reasoning preflights passed, as did a 27/27 coding-agent slice and a warmed 125,785-prompt-token request. Against the same image and checkpoint without speculative decoding, DSpark improved median per-request decode by 101.4% and reduced median E2E by 58.8%. It also consumed 1.6-2.3 GiB more VRAM, and neither lane retained 3 GiB reported free on each card. See the r16 qualification.

Current 0731-specific NVFP4 conversions are community artifacts. The strongest conversion receipt reports TP=2 generation on two DGX Spark systems; the strongest RTX PRO performance report uses four cards and an out-of-tree DSpark path. Neither proves fit, quality, or speed on this two-card topology. See the research update for pinned identities, benchmark deltas, conversion recipes, GGUF sizes, and the publisher-reasoning/DSpark/NVFP4 qualification sequence.

Single-card and historical profiles

Candidate Repeated quality Context evidence Decision
Qwen3.8 27B 2026-09-04 SGLang comprehensive campaign TP1/DP2, RadixArk, and kelnei/vLLM passed smoke/JSON/tools/Responses; all 100-request headline cells passed 100/100 canaries; TP2 failed JSON twice sustained-output aggregate: TP1 764.3, TP2 587.9, DP2 1,401.8–1,423.4 tok/s; RadixArk 746.7 with 46.9% lower median TTFT but 6.8% higher median E2E; kelnei MTP2 503.4 versus 315.2 no-spec; unique 82K/C8 median E2E 93.7-137.6 s DP2 bounded throughput winner, TP2 rejected, all no-promotion; balancing/failover, broad quality, clients, multimodal, and power/energy remain open; exact baseline restored
Qwen3.8 27B SGLang official FP8, remote AI-MBP25 worker agentic smoke 2/2; agentic scout 16/18 with both failures in debug-loop; fixed SWE-bench Verified scout 5/5 officially graded and resolved 393,216-token service; SWE tasks required 19-57 model requests and completed in 42m29s overall bounded coding-agent evidence from the former service; larger stratified SWE and efficiency-aware debugging follow-up
Qwen3.8 27B SGLang official BF16 / official FP8 / Inferact NVFP4, MTP=3 multimodal All pass the same earlier 18/18 image corpus; the official FP8 arm later passed direct 30/30 including video 14/14 and live admitted 28/28 393,216 configured tokens; at 4K BF16/FP8/NVFP4 measure 62.7/111.4/97.7 decode tok/s; official-FP8 direct video p50/p95 is 2.935/9.904 s Official FP8 is the former human-promoted single service; NVFP4 is the lowest-latency third-party no-promotion alternative with video still open
Qwen3.8 27B official BF16 / official FP8 vLLM split Both passed intelligence/session/tools at 3/3 and adaptive low/medium/xhigh control; final routed BF16 media 30/30 and 32-image request 1/1; all 16 TP/MTP matrix arms passed complete functional gates Split TP=1 and exclusive TP=2 passed 388,979 actual tokens; TP=2 also passed 598,729 and 985,107 on both checkpoints, with 13.0-13.7 minute near-1M TTFT managed rollback at TP=1/393K/MTP=3; TP=2 remains batch-like
Qwen3.8 27B SGLang official FP8 / Inferact NVFP4, MTP=3 Both pass intelligence 6/6, session 3/3, tools 3/3, plus CPU-transport image/OCR; the later corpus adds repeated two-image ordering Both pass 389K retrieval; at 4K official FP8 averages 111.3 decode tok/s while NVFP4 averages 98.1, with NVFP4 retaining lower TTFT and higher prefill official FP8 former deployment; Inferact NVFP4 no-promotion
Qwen3.8 27B SGLang official FP8 / Inferact NVFP4 Both pass full functional gates on two card placements plus intelligence 6/6, session 3/3, and tools 3/3; text-only, no speculation Both pass 388,979 actual tokens; NVFP4 248.75 s TTFT / 1,564 prefill tok/s versus official FP8 258.13 s / 1,507; at 4K NVFP4 averages 57.9 versus 48.0 decode tok/s across five runs NVFP4 text challenger; both no-promotion; former vLLM MTP=3 was faster and SGLang multimodal was unqualified on WSL2
Qwen3.8 27B official FP8 MTP=4/5 Both pass tools 20/20 and repeated deterministic intelligence/session/tools; the quality artifacts are behavioral rather than timing evidence Both pass one cold 388,979-token request; cross-card 4K runs show MTP=5 only 0.4-1.3% above MTP=4 decode on a fixed card and no E2E win MTP=4/5 no-promotion; MTP=3 remained the selected Qwen depth
Qwen3.5 122B NVFP4 Passed protocol-v3, tools 10/10, image/OCR 128K and 240K retrieval; 262,144 served rollback
Laguna S 2.1 NVFP4 Passed protocol-v3 with thinking disabled 32K/128K/240K passed; TTFT 2.26/21.15/50.64 s rollback
GPT-OSS Puzzle 88B Tools/session/timeout 3/3; unified diff 2/3 32K and 128K retained rollback
Agents-A1 Official FP8 passed three-repetition protocol-v3; BF16/FP8 retain the identical 28/30 multimodal corpus 240K on a 262,144 serve; 35.21 s promotion-quality TTFT, 32.97 s capacity TTFT p50 at 231,426 actual tokens historical promotion
Gemma 4 12B QAT W4A16 Historical protocol-v3 control passed 240K passed no-promotion
ThinkingCap Qwen3.6 27B FP8 Historical strict-quality control passed 240K retained no-promotion

Capacity and recipe comparisons

Configured concurrency is not full-window capacity. In the GLM rows below, maxseq16 means the scheduler may run sixteen requests; it does not mean the KV pool can hold sixteen requests at the served context ceiling. The historical 393K SGLang profile was configured and qualified at C1. The corrected 524K rollback proved C2 at 206,630 prompt tokens/request, while the historical BrandonMusic C16 artifacts used 4K prompts. See the cross-run capacity interpretation.

Model/configuration Served context Admission Capacity note
GLM-5.3-Flash EXL3 r7 no-spec, TP=2 327,680 C4; text only Selected 2026-09-14; 65,536 output reserve; 9/9 context through 255,647–255,672 prompt tokens; 1,638,400 reported shared KV tokens is allocation, not proof of simultaneous full-window correctness
GLM-5.3-Flash ormandj W4A16/NVFP4, SGLang adaptive EAGLE, TP=2 393,216 1; image/OCR; video zero historical 2026-09-02 selection; FP8 KV; 304,491 actual prompt tokens at the 380K target; 2,101 MiB/card after qualification and 2,543 after client work under an explicit model-only zero-reserve waiver; 240K remains the conservative verified fallback
GLM-5.3-Flash EXL3 K3 plus corrected DFlash2 K5, TP=2 524,288 router 16; measured C2 at nominal 250K; up to 16 images, no video historical rollback; FP8 DS-MLA target KV and BF16 draft KV; 2,493,817 KV tokens / 4.76 complete windows; 206,296-actual-token retrieval; C2 2/2; batch4,096 rejected
Qwen3.8 Flash Next RadixArk NVFP4, SGLang QSA-fast MTP3, TP=2 262,144 1; four images or one video historical video-capable rollback; 516,032 maximum server tokens; 6.275 GiB KV per rank / 12.55 GiB aggregate; 253,703 actual prompt tokens with an 8,192-token output request; direct media 30/30; BF16 KV, no PLE CPU offload; exact PR #36556 SM120 QSA path; matched NEXTN 3/1/4 qualification
GLM-5.3-Flash TR3/EXL3 4 bpw, vision fixed K5, TP=2 262,144 16 configured; c1 long-context measured; one image proven NVFP4 DS-MLA KV; reports 560,866 KV tokens / 2.14 configured windows, measures 72.8/55.7 tok/s decode at 4K/128K, passes a 250K target / 206,296 actual prompt-token retrieval, and completes c16 short 16/16 at 28.3 aggregate output tok/s; image/OCR pass, video disabled
GLM-5.3-Flash TR3/EXL3 4 bpw, text fixed K5, TP=2 262,144 / 524,288 16 configured; c1 long-context measured NVFP4 DS-MLA KV; 262K lane reports 688,890 KV tokens and measures 69.8/61.9 tok/s decode at 4K/128K; 524K lane reports 565,898 KV tokens, passes 495,045-token retrieval and 497,976-token tool use, and is single-user due to the 1.08x KV envelope
GLM-5.3-Flash TR3/EXL3 4 bpw, no spec, TP=2 262,144 / 524,288 16 configured; c1 long-context measured NVFP4 DS-MLA KV; 262K lane reports 1,775,814 KV tokens and completes c16 short 16/16; 524K lane reports 1,603,111 KV tokens, passes 495,045-token retrieval and 497,976-token tool use, and is the preferred maximum-context/headroom profile
DeepSeek V4 Flash 0731 Infernal Invocation r18 B12X + DSpark K5, batch 4,096, TP=2 1,048,576 8 engine; router 1 FP8 compressed MLA KV, GPU-only; 1,323,176 reported KV tokens; 1,040,063 actual prompt tokens pass; c8 short and c2 at 490,861 tokens/request pass; matched no-spec control retained; former Primary
DeepSeek V4 Flash 0731 Infernal Invocation r15 B12X + DSpark K5, batch 4,096, TP=2 393,216 8 engine; router 2 FP8 compressed MLA KV, GPU-only; 797,689 reported KV tokens; 351,118 direct and 340,119 routed actual prompt tokens passed; c8 short and c2 at 99,175 tokens/request passed; matched no-spec control retained
Qwen3.8 27B 2026-09-04 SGLang Inferact/RadixArk and kelnei/vLLM 262,144 C8 per TP1/TP2 service; aggregate C16 across two TP1 replicas SGLang K12/chunk1K/Mamba96 selected; TP1/TP2/DP2 and RadixArk measured; kelnei vLLM MTP2/no-spec measured; all unique 82K/C8 arms completed but were non-interactive; TP2 rejected for speed and strict JSON
Qwen3.8 27B SGLang official BF16 / official FP8 / Inferact NVFP4, TP=1 MTP=3 multimodal 393,216 1 each All pass the earlier 18/18 image corpus; official FP8 later passed video 14/14 and live admitted 28/28 at two images/one video; NVFP4 video plus broader concurrency, memory-pressure, and quality gates remain open
Qwen3.8 27B SGLang official FP8 / Inferact NVFP4, TP=1 MTP=3 393,216 1 each Both pass 389K retrieval and cross-card gates; official FP8 averages 111.3 decode tok/s, NVFP4 98.1 with 0.448 s TTFT; CPU transport passes bounded image/OCR on both but full media and host-memory-pressure gates remain open
Qwen3.8 27B SGLang official FP8 / Inferact NVFP4, TP=1 no-spec 393,216 1 each Both pass 388,979 actual prompt tokens and cross-card functional gates; official FP8 reports 1,665,740 KV tokens and 48.0 tok/s 4K decode, NVFP4 1,805,068 and 57.9 tok/s; text-only because WSL2 CUDA-IPC multimodal warmup failed
Qwen3.8 27B official FP8 text, TP=2 control/MTP=3 393,216 / 600,000 / 1,010,000 1 All three limits passed at 388,979 / 598,729 / 985,107 actual tokens; control reports ~4.65-4.67M KV tokens; MTP reports ~4.22-4.33M and 85.9-91.6 tok/s 4K decode; no P2P
Qwen3.8 27B official BF16 multimodal, TP=2 control/MTP=3 393,216 / 600,000 / 1,010,000 1 All three limits passed; control reports ~3.77-3.78M KV tokens; MTP reports ~3.42-3.50M and 67.4-75.6 tok/s 4K decode; no P2P
Qwen3.8 27B official FP8 text 262,144 control / 1,010,000 continuation 5 control / 1 continuation Control reports 1,825,809 KV tokens and 51 aggregate output tok/s at c5; 1M arm reports 1,845,432 KV tokens and passes 825,049 actual prompt tokens 3/3, but cold E2E is ~956.7 s
Qwen3.8 27B official BF16 multimodal 262,144 2 FP8 KV reports 1,036,311 tokens / 3.95 full windows; 241,250-token retrieval and 30/30 media corpus pass
DeepSeek V4 Flash 0731 r33 B12X + DSpark K5, batch 4,096, TP=2 393,216 16 configured; c1 long-context measured FP8 DS-MLA KV, GPU-only; 725,543 reported KV tokens; 359,900 actual prompt tokens passed; managed exclusive rollback
DeepSeek V4 Flash 0731 r33 B12X target-only, batch 4,096, TP=2 131,072 measured; GPU-only 393,216 next 1 FP8 DS-MLA KV; 119,503 actual prompt tokens passed; lowering batch tokens reduced activation 34.1% and increased minimum-rank KV 15.27 to 15.99 GiB; engine reports 553,243 tokens, but >300K remains unproven pending a configured long-context request
DeepSeek V4 Flash 0731 r33 B12X target-only, TP=2 131,072 measured; 393,216 translated only 1 FP8 DS-MLA KV; 119,503 actual prompt tokens at 73.86 decode tok/s; 283,917-token GPU KV pool cannot support 393K without host capacity; translated recipe adds 16 GiB native offload
Agents-A1 FP8 multimodal 262,144 c1 at 262K; earlier 131K c32 188 tok/s decode at 8K c1; 156 tok/s decode and 32.97 s TTFT at 240K; 51.93 GiB KV; generated MoE tune rejected
Agents-A1 NVFP4 compact text 131,072 16 198 tok/s at 8K c16; 128K c4 pass; vision excluded
Qwen3.5 122B NVFP4 262,144 1 BF16 KV; near-ceiling prefill is slow
Laguna S 2.1 NVFP4 262,144 recorded recipe FP8 KV; disabled-thinking contract
DeepSeek V4 Flash 0731 r16 B12X + DSpark K5, TP=2 131,072 8 configured; c1 measured FP8 MLA KV; 130.7 tok/s matched decode; 128K pass; 3 GiB reserve failed
DeepSeek V4 Flash 0731 r16 B12X + DSpark K5 GPU-only Pi, TP=2 650,000 historical / 1,000,000 experimental 16 650K: 640K retrieval, 141.6 tok/s matched 32K decode, live Pi/OpenClaw smokes; 1M: retained client-shaped workspace failures; reserve waived
DeepSeek V4 Flash 0731 r16 B12X + DSpark K5 + native offload, TP=2 262,144 8 configured; c1 measured 8 GiB cold 250K capacity; 16 GiB CPU tier proves 113,408-token external reload; per-card reserve not sampled
DeepSeek 0731 Vision (NVFP4), webbrain-one SGLang, TP=2 4,096 1; one image per request ~175.6 GB mixed FP8/NVFP4 weights, 88.08/87.87 GB per card; --mem-fraction-static 0.97 against an engine-measured KV floor of 0.9411; steady-state 95,164/93,992 MiB of 97,887 MiB; marlin MoE JIT first-compile requires a persistent tvm-ffi cache volume
GPT-OSS Puzzle 88B MXFP4 131,072 8 FP8 KV; pinned Anvil vLLM
Nemotron 3 Super 120B NVFP4 131,072 5 1M advertised is not locally validated
Qwen3.6 27B community NVFP4 + MTP 262,144 5 262K needle validated
Mistral Small 4 119B NVFP4 131,072 5 Low short-request TTFT; weaker quality slice

External recipe watch and local follow-up

The 2026-08-15 Qwen3.8 recipe refresh recorded two test-next directions. The first is now complete: the MTP-depth qualification found no meaningful MTP=4/5 E2E win against the then-current official-FP8 TP=1/393K MTP=3 lane. SGLang's commit-pinned cookbook supplies explicit RTX PRO 6000 official-FP8/BF16 cells, SM120 FlashInfer guidance, 2,048-token prefill chunks, and GDN state-cache sizing controls. The SGLang 200+ tok/s headline uses third-party NVFP4 and DSpark artifacts and is not comparable with the local official-weight result.

Dormant vLLM MTP=4/5 recipes remain measured no-promotion controls. The SGLang official-FP8/NVFP4 qualification completed the second direction with a digest-pinned runtime, exact revisions, cross-card measurements, and exact restoration. NVFP4 improved the matched no-spec SGLang lane. The later MTP/multimodal qualification raised official-FP8 decode to 111.3 tok/s and NVFP4 to 98.1, while bounded image/OCR passed on both with CPU feature transport. The official-FP8 arm was subsequently human-promoted as the single-service Primary/general-vision/OCR profile and then qualified for one video; Inferact NVFP4 remains no-promotion.

Recent changes

  • 2026-09-08: native Linux / retained WSL GLM comparison added above; same measured dual-card pair, faster historical-style decode with explicit failed strict-output and routed gates; no new promotion.

  • 2026-09-04: the current-runtime Qwen3.8 27B campaign optimized SGLang through draft depth, chunking, compile, Mamba allocation, topology, and target checkpoint rounds, then ran kelnei/vLLM MTP2/no-spec. On the matched 100-request sustained-output workload, one TP1 measured 764.3 aggregate tok/s, TP2 587.9, and two TP1 replicas 1,401.8–1,423.4 at aggregate C16. TP2 also failed strict JSON twice. RadixArk is the lower-TTFT tradeoff; kelnei MTP2 improved 59.7% over no-spec. Unique 82K/C8 remained non-interactive and broader quality/power/client gates are open. The exact GLM service, route, GPU ownership, and clean shared memory were restored; no promotion occurred.

  • 2026-09-03: a cross-run GLM review separated scheduler ceilings, shared/reported KV-token pools, short-request batching, and measured long-context concurrency. The current 393K SGLang profile remains C1; the immediate corrected 524K rollback retains a measured C2 206,630-token/request gate; historical BrandonMusic maxseq16 results remain short-request evidence. No model was loaded and no route, recipe, or deployment changed.
  • 2026-09-02: the pinned ormandj GLM-5.3-Flash W4A16/NVFP4 rc14 SGLang profile was translated to WSL2, qualified through 304,491 actual prompt tokens at TP=2/393,216/C1, and human-promoted under an explicit model-only reserve waiver. Direct, managed, routed, and real Pi/OpenClaw/Hermes gates passed; post-client free VRAM was 2,543 MiB/card with no OOM, restart, crash, CUDA error, or shared-memory residue. The first promotion attempt exposed a missing media-fixture manifest contract, which was fixed and tested before the successful retry. The exact 524K profile is retained as rollback.
  • 2026-09-02: the GLM 393K subject resolved the fixed django__django-11099 SWE-bench Verified smoke 1/1 under the official grader from an isolated macOS worker. The run used 11 routed requests and completed its agent/grader stages in 34.216/29.076 seconds. Revision-bound harness dependencies and ambient worker configuration were fixed forward before the retained run; the result is bounded infrastructure and repository-agent evidence, not a full-suite score.
  • 2026-08-31: a digest-pinned xgrammar correction removed the DFlash2 structured-generation failure, and the matched 524K K5 arm improved decode from 42.61 to 83.08 tok/s at 4K and from 43.63 to a pooled 69.99 at 240K. The engine reported 2,493,817 KV tokens, C2 nominal 250K completed 2/2, and direct, rollback, router, OpenClaw, Hermes, and Pi forward gates passed. The corrected profile is current; the former 1M profile is the first rollback.
  • 2026-08-30: the exact GLM-5.3-Flash EXL3 K3 target plus DFlash2 K5 draft became the human-authorized one-week text/image/OCR default at TP=2/1,048,576/maxseq16/router-c16. It measured 82.1/67.4/67.9 tok/s at 4K/131K/240K, passed exact 950K-target retrieval, tools 20/20, image/OCR 12/12, bounded quality 12/12, routed identity, and real Hermes/Pi/OpenClaw acceptance. K3 remains a verified alternate; batch4,096 is rejected. Video remains unsupported and the DFlash2 draft is noncommercial without separate permission. Qwen3.8 Flash Next is the immediate video-capable rollback.
  • 2026-08-29: the exact Cardillo/Purtell GLM-5.3-Flash TR3/EXL3 4 bpw checkpoint/runtime pair qualified under WSL2 at TP=2/DCP=2. Fixed K5 passed vision/OCR, tools 20/20, a bounded high-reasoning coding suite 15/15, and a 250K-target / 206,296-actual retrieval while measuring 72.8/55.7 tok/s decode at 4K/128K. Both 524K text profiles passed exact retrieval at 495,045 actual prompt tokens and valid tool use at 497,976; no-speculation retained 1,603,111 reported KV tokens. Adaptive K1-K5 plus ReplaySSM was rejected for repeated tool corruption, and 0xSero 3.0-bpw remains watch-only. The candidate remains no-promotion, the current Qwen route is unchanged, and the direct vision service is running for hands-on.
  • 2026-08-26: the exact RadixArk Qwen3.8 Flash Next NVFP4 revision and digest-pinned SGLang image qualified at exclusive TP=2/262,144/c1, then fixed forward to the hash-gated PR #36556 SM120 QSA fast path and matched MTP3 3/1/4. MTP3 passed tools 20/20, bounded quality 12/12, the full-reserve request, exact routed identity, and fresh OpenClaw/Hermes/Pi acceptance. It measured 154.9 tok/s at 4K, 134.1 at 128K, and 102.0 at full reserve, without a bounded quality regression. A same-service vision campaign then passed direct media 30/30, live routed repeats 57/60 strict, admission/SSE/tool/error edges 8/8, and 25/25 requests across a six-size context curve through 245,000 actual prompt tokens. The route now owns image/OCR/video at four images or one video; the portable 12.8 tok/s lane remains a superseded correctness baseline.
  • 2026-08-21: the exact digest-pinned Infernal Invocation r18 K5 profile qualified at 1,048,576 tokens with 1,040,063-token retrieval, complete functional and post-reload gates, agentic 12/12, additional tools 160/160, client-shaped reserves, c8 short, c2 long, and a matched no-spec A/B. The operator-authorized guarded promotion, exact routed identity/API/tool gate, Mini generation-2 convergence, and real Hermes/Pi/OpenClaw acceptance passed. r18 was the current text Primary until the 2026-08-26 Qwen promotion.
  • 2026-08-16: after explicit human approval, the digest-pinned DeepSeek Infernal Invocation r15 K5 profile became the exclusive TP=2 text Primary at 393,216 tokens. Matched K5/no-spec performance, 351,118-token direct and 340,119-token routed retrieval, repeated quality 12/12, c8 short, c2 long, tools, streaming, Responses, exact managed routing, and rollback checks passed. The r33 393K profile is the fixed-port managed rollback. Martin Vit's upstream receipt covered 131,072 tokens on native Linux with two RTX PRO 6000 Blackwell GPUs on direct PCIe root ports; the 393K WSL2 result is independently qualified. Actual Mini OpenClaw remains open because the installed Mini controller lacks the current status tool.
  • 2026-08-16: the then-current official-FP8 SGLang service passed direct media 30/30 with video 14/14. A managed router-only expansion added vision.video and fail-closed two-image/one-video admission; live admitted media passed 28/28 plus overflow, malformed-input, SSE, tool, and Primary regression gates. The model was not restarted and the second card remained dormant.
  • 2026-08-15: the exact official-FP8 SGLang TP=1/393K/MTP 3/1/4 profile was human-promoted on one card, with the other card left empty. Guarded 108K and 20-tool checks, direct+routed 18/18 media, routed Responses, and fresh Hermes/OpenClaw Primary turns passed. Initial admission was one request, two images, and no video; the former vLLM FP8/BF16 split remains rollback.
  • 2026-08-15: a matched SGLang consolidation A/B added official BF16 to the MTP=3 CPU-transport comparison and ran 18 repeated media attempts per model, including two-image ordering. All three passed. Official FP8 cut media p50 35.8% and raised 4K decode 77.7% versus BF16; NVFP4 cut media p50 51.1% and halved TTFT but decoded 12.3% slower than official FP8. Official FP8 is the preferred single-service challenger; the current split was restored and routed, with no promotion.
  • 2026-08-15: SGLang EAGLE MTP 3/1/4 raised matched official-FP8 decode from 48.0 to 111.3 tok/s and Inferact NVFP4 from 57.9 to 98.1. Both passed 389K retrieval, repeated deterministic quality, and bounded image/OCR after forcing CPU feature transport around the failing CUDA-IPC path. The exact vLLM split was restored and readmitted; no route or promotion changed.
  • 2026-08-15: digest-pinned SGLang official FP8 and audited Inferact NVFP4 both passed full functional, repeated deterministic-quality, cross-card 4K, and 388,979-token gates at TP=1/393K. NVFP4 averaged 22.6% lower TTFT and 20.6% higher decode than the official SGLang control, but remained well behind current vLLM MTP=3 decode. The exact current split was restored and readmitted; both SGLang recipes remain no-promotion and text-only.
  • 2026-08-15: official-FP8 MTP=4 and MTP=5 both passed functional, deterministic-quality, and 388,979-token gates. A cross-card swap showed the first-placement speed gap followed the GPU lane; MTP=5 beat MTP=4 decode by only 0.4-1.3% on the same card and did not improve E2E. The exact MTP=3 FP8/BF16 split was restored and readmitted; MTP=4/5 remain no-promotion.
  • 2026-08-15: external research queued a matched official-FP8 MTP=3/4/5 vLLM A/B (now completed above) and an official-weight SGLang compatibility spike. Third-party NVFP4, DSpark, GGUF, AutoRound, and custom-runtime artifacts remain excluded. This is an external-prior update, not a new hardware measurement.
  • 2026-08-14: human approval promoted the matched 393K TP=1/MTP=3 split. Official FP8 is the text Primary and official BF16 handles explicit general-vision/OCR with a 32-image request ceiling. Routed FP8 functional gates, BF16 media 30/30, one 32-image request, and Hermes/OpenClaw client paths passed without fallback.
  • 2026-08-14: the matched Qwen3.8 BF16/official-FP8 matrix completed split TP=1 at 393K and exclusive TP=2 at 393K/600K/1.01M, each with control and MTP=3. All 16 arms passed full functional gates and cold retrieval at 388,979/598,729/985,107 actual prompt tokens. MTP raised 4K decode 1.76-2.40x while consuming 7-11% of reported KV tokens; TP=2 cut 393K control TTFT 35-38%. The exact 262K split baselines were restored; no route changed.
  • 2026-08-14: the official Qwen3.8 27B FP8 checkpoint was configured for 1,010,000 tokens on one card and passed retrieval through 825,049 actual prompt tokens at 3/3. Mean cold request-to-completion latency was 956.739 seconds, so the result is stable offline/batch capacity rather than an interactive recommendation. The post-stress gate passed and the exact 262K FP8 lane was restored; decision remains challenger, no-promotion.
  • 2026-08-14: official Qwen3.8 27B BF16 multimodal and official FP8 text checkpoints qualified on concurrent one-card split lanes. Both passed the functional, adaptive-reasoning, repeated-quality, 4K capacity, and 241,250 actual-prompt-token gates; BF16 passed 30/30 media attempts. FP8 MTP=3 raised c1 decode 47.9 to 94.8 tok/s, prefix caching reduced a repeated 30K-prefix burst to 0.41 s TTFT, and unquantized KV halved full-window capacity without a 4K speed gain. Decision challenger, no-promotion; no route changed.
  • 2026-08-11: after human approval, the r33 DSpark K5 GPU-only profile became llm.primary at 393,216 tokens, maxseq16, and batch 4,096. Direct capacity passed through 359,900 actual prompt tokens; OpenClaw and Hermes were aligned to 393,216 context, 32,768 output, and high reasoning, then passed client-path requests after gateway restarts. The legacy routed nominal-320K needle failed 413 because its conservative byte estimate exceeded the route limit, and the SWE smoke did not submit because installed benchmark profiles were missing. Both limitations remain explicit.
  • 2026-08-10: a matched r33 A/B halved max_num_batched_tokens from 8,192 to 4,096. Fresh bracketed starts reproduced the 8,192 baseline and measured a 34.1% activation reduction plus a 0.72 GiB minimum-rank KV increase. The 4,096 arm passed the same 6/6 functional and 119,503-prompt-token capacity gates and was healthy/direct-only at campaign close. Its 553,243 reported KV tokens are not treated as >300K proof because KV bytes increased only 4.715%; the next gate is a GPU-only 393K configured serve and actual >300K request.
  • 2026-08-10: the digest-pinned r33 B12X target-only control passed the full functional preflight, repeated high-reasoning intelligence/session/tool checks, and a 119,503-prompt-token request at 73.86 decode tok/s. The engine exposed 283,917 GPU KV tokens, below the 393K target; a quality-first 393K recipe retains FP8 DS-MLA KV and adds 16 GiB native host offload but was not loaded. Requested context targets were non-monotonic versus API-reported prompt tokens, so a benchmark-integrity ticket remains open. No route or promotion changed.
  • 2026-08-07: a WebBrain DeepSeek 0731 vision-adapter (NVFP4) package first-loaded and served on TP=2 via SGLang, marlin/marlin kernels, and --mem-fraction-static 0.97 against an engine-measured KV floor of 0.9411. ~175.6 GB of mixed FP8/NVFP4 weights split 88.08/87.87 GB per card, with steady-state usage of 95,164/93,992 MiB of 97,887 MiB; the first marlin MoE call required a persistent tvm-ffi JIT cache volume. Image conditioning was grounded, but OCR/GUI reading confabulated and the checkpoint has no chat template. Decision no-promotion; the 650K Primary was restored and verified healthy in the same session.
  • 2026-08-03: AI-MBP25 completed the first managed remote context, agentic recovery, and SWE-bench Verified smoke against the unchanged 650K DeepSeek Primary. The 8K context and one-instance official grader paths passed. The tool-error case retried correctly but failed its final answer. This qualifies the worker and artifact path for a scout campaign without changing routes.
  • 2026-08-02: after human approval, the 650K/maxseq16 profile became llm.primary with high reasoning as the client default and a generic per-tier 32,768 output cap. Dark Pi, Mini Pi, and Mini OpenClaw passed. The 1M/maxseq16 profile was removed after two real client-shaped B12X workspace crashes, including one with only 5,120 requested output tokens.
  • 2026-08-02: moving display output to the AMD iGPU allowed the DeepSeek 0731 maxseq16 envelope to start. GPU-only 650K/maxseq16 and 1M/maxseq4/maxseq16 passed near-limit retrieval and Pi protocol gates. The 650K profile is the preferred everyday Pi experiment; 1M/maxseq16 is the preferred explicit deep-session alternative. A retained 1M/maxseq1 B12X workspace crash and sub-1-GiB post-workload free VRAM keep all profiles no-promotion.
  • 2026-08-02: the derived DeepSeek 0731 WSL2 native-offload lane qualified a 262,144-token serve through a 249,573-prompt-token request. A 16 GiB CPU tier also reloaded a 113,408-token external prefix after GPU eviction. An ownership-aware Anvil lifecycle now blocks cleanup while workers map the file and reclaims both 8 and 16 GiB mmaps after teardown. The decision remains no-promotion.
  • 2026-08-01: the pinned DeepSeek 0731 r16 B12X lane qualified DSpark K5, low/high/max reasoning, 128K, and 27/27 coding-agent attempts. DSpark doubled matched decode versus same-image no-spec, but both lanes failed the 3 GiB reserve and remain no-promotion.
  • 2026-08-01: DeepSeek V4 Flash 0731 became the priority intelligence challenger after official and independent research was reconciled with its exact local low-reasoning TP=2 evidence. The decision remains no-promotion; no new GPU run or route change occurred in the research pass.
  • 2026-08-01: the hardware became a symmetric two-PRO topology. The exclusive TP=2 campaign qualified Qwen3.5, Nemotron 3 Super, Laguna S, DeepSeek V4 Flash 0731, and Inkling Small without changing production aliases.
  • 2026-07-29: Agents-A1 official FP8 passed the missing three-repetition protocol-v3 suite at the 262K profile and was promoted through the managed transaction. Qwen3.5 is now the immediate rollback.
  • 2026-07-29: Agents-A1 official FP8 passed the same 262K/240K functional and capacity shape as Qwen. It used 35.31 versus 73.22 GiB model memory, halved 240K TTFT, and delivered the unchanged video corpus. Qwen passed all images but its exact NGC image lacked an H.264 decoder. Agents-A1 wins the bounded comparison; Qwen remains Primary pending matched repeated quality.
  • 2026-07-28: Agents-A1 BF16/FP8 image and direct-video capability passed, but both reached only 28/30 on the strict multimodal corpus. NVFP4 qualified as a compact text-only profile. Isolated routed video passed after bounded thinking/error-classification fixes, and the FP8 tune was rejected; no route changed.
  • 2026-07-28: Qwen3.5 122B became the human-gated Primary; Laguna moved to immediate rollback.
  • 2026-07-27: Agents-A1 qualified as a thinking-disabled challenger without promotion.
  • 2026-07-26: Laguna S 2.1 passed repeated quality and 240K retrieval.
  • 2026-07-18: Puzzle established the pinned secondary recipe and strict-format caveat.

Run history

The complete PRO 6000 history, including failed loads and incomplete runs, is in the run catalog. Every row links its dated finding and a stable model dossier.