Skip to content

Qwen3.8 Flash Next

2026-10-26 three-recipe scout — final, no-promotion: distinct Ignis 0.8.0, vLLM r2, and Flare r2 recipes each passed the six-case agentic suite and nine-sample context scout (maximum observed prompt 114,588). Aider and official SWE results were mixed; the shared router's buffered-text whitespace defect makes historical coding attribution uncertain because original HTTP bodies were not retained. Flare C1 ended 0/8 performance-eligible; its native native_control_mismatch is preserved. Exact prior evaluation route/output-cap configuration and GLM incumbent readmission passed local checks; direct/routed five-check preflights and one Pi real read-tool turn passed. The router uses reviewed PR #704 overlay, not original image. No engine ranking, speed result, full-context/capacity qualification, or promotion follows. Finding · evidence.

2026-10-26 Ignis 0.8 mandatory-tool qualification: the original one-card tool_choice: "auto" preflight retained a 3/4 tools batch. A separately frozen same-runtime follow-up used the documented request-level required policy and normal thinking: 3/3 mechanism cells, 5/5 selected checks with 9/9 observations, and a seed-18 4/4 tools repeat passed. All 16 phase-two requests have terminal engine events; literal HTTP 200 was retained for 12 captured tool requests. Exact operational restoration passed after both phases. This qualifies the mandatory-tool request path on the exact pin, not auto reliability, a global agent default, coding/SWE, capacity, performance, clients, or promotion. Finding · evidence.

2026-10-16 NVIDIA FP4 scout: the pinned Qwen profile passed native preflight 6/6 and Oracle 54/54, then ended the five-case SWE run at 2/5 resolved and 4/5 graded. It missed the frozen 4/5 floor; vision, context, capacity, and stability are deliberate non-results. Finding

Current status and review date

2026-10-25 R12/R13 successor: R13 froze a fresh 48-cell Ignis 0.8 candidate-only attempt but submitted zero jobs: the host-reserve guard refused before container creation. This is separate from R12's 48 unrun candidate cells and the earlier R9/R10 attempts. Exact incumbent restoration and explicit minimal-thinking Pi acceptance passed; historical thinking-off failure remains open. No candidate qualification, comparison or promotion follows. Finding · Authoritative summary.

Historical R9/R10 Ignis 0.8 addendum, 2026-10-25: R9/R10 admission diagnostics produced no generation or coding result. R9 was refused by the host-reserve gate; R10 loaded but Anvil Serving withheld monitoring after misreading its served model identity. Both 48-row reports remain entirely planned. The incumbent service was restored; strict Pi acceptance and a fresh post-fix candidate campaign remain required. No qualification, ranking or promotion follows.

2026-10-19 addendum: Ignis0.7.2 is not qualified. Bounded serving checks passed, but the fixed-five coding run ended with one officially unresolved patch and four empty ungraded step-limit submissions. The incumbent was restored; no recommendation change. Evidence and limits.

2026-10-05 replacement scout: the evaluated prospective primary profile resolved 3/5 frozen coding tasks against a 4/5 floor and its strict C4 128-word cell was 0/4 performance-eligible. It is rejected with no-promotion. These results apply to that profile only and do not revise the historical Qwen evidence below. See the campaign finding.

October 5: separate 5090 expert-offload investigation

The 96 GB host campaign measured Strata UD-IQ4_XS on one RTX 5090. Initial 32K/C1, R36/M50 passed five protocol families and 12/12 image checks. All twelve service canaries passed, but native nearest-rank p50 visible TTFT 30.406 s and E2E 30.852 s failed the frozen practical latency gate against Swift. Three-request instrumented MMQ/fused diagnostics are separate populations, not replacement gates. A final fixed-4096 diagnostic shifted refill cost into batched prefill without useful overall improvement. This configuration was rejected for promotion; larger-context/C4 and finalist quality gates were not run. No local quality advantage is established; Swift host-cache profile subsequently passed final acceptance and was promoted. Review: 2026-10-05; earlier lanes below retain their own hardware, quantization and decisions.

Decision snapshot

  • Product role: measured EXL3 alternative to the September 14 GLM text selection; separate NVFP4 multimodal promotion history.
  • Selected or best-qualified configuration: EXL3 4.05-bpw at 262,144 configured tokens in the September scout. The August NVFP4/SGLang TP=2/C1 recipe has separate text, image, OCR, and video acceptance.
  • Measured hardware: two RTX PRO 6000 Blackwell Max-Q cards; native Linux for EXL3, WSL2 for the historical NVFP4 lane.
  • Evidence: EXL3 scored 91/100 on one fixed MMLU-Pro sample, context 9/9, agentic 21/30, SWE 4/5, and image/OCR 12/12.
  • Decision: EXL3 remains no-promotion; GLM no-spec was selected in that campaign. NVFP4's August rollback role is historical.
  • Important limitation: EXL3 strict120 has zero performance-eligible responses because leading line feeds violated the canary contract. Neither its timings nor a one-point quality gap establish superiority.
  • Review dates: retained evidence through 2026-09-22; dossier reviewed 2026-09-22.

Latest finding · raw evidence · configuration history.

Review narrative

2026-09-23 — pinned NVFP4 candidate did not pass strict capacity

The medium profile resolved 4/5 frozen SWE tasks and passed Oracle7 18/18, but its strict C4 repair population was 0/4, so it is performance-ineligible and unqualified. Separate default-xhigh diagnostics passed preflight 26/26, vision 4/4, and a 412,589-token retrieval; those results do not qualify medium. No matched no-MTP control or controlled speed result exists. Evidence.

2026-09-22 — NVFP4 TP2 with C4 configured admission; serial C1 replacement scout did not advance

The distinct RadixArk NVFP4 recipe used TP2 with C4 configured admission; its direct and agentic checks were serial C1. It loaded on native Linux and passed six direct preflight groups. Its native agentic results were 16/18 greedy and 15/18 with official sampling. Those artifacts pass their internal 0.75 floor, but the campaign requires 18/18 deterministic success, so this candidate did not advance to context, capacity, performance, vision, deep-coding, or SWE work. Both greedy misses violate exact tool-call count/arguments; positional fixture drift then limits interpretation. The sampling replay had three misses. A paired GLM 16/18 scout exposed a planning-scorer false negative, so these artifacts cannot rank broad coding or show Qwen is worse. Outcome: no-promotion; retain GLM while harness validity blocks heavy qualification.

The startup log retains a nonfatal custom-allreduce UUID-to-index parse warning and fallback; readiness and direct preflight still passed, so it is a runtime caveat rather than a clean-start claim.

2026-09-13–14 — EXL3 alternative retains quality and formatting gaps

The 4.05-bpw EXL3 profile completed the fixed quality sample, long-context, agentic, SWE, and image/OCR probes. GLM no-spec completed 30/30 agentic tasks against Next's 21/30; both resolved 4/5 frozen SWE tasks. These are bounded samples across different configurations, not a general model ranking. All 120 Next capacity responses stopped, but none met the strict canary-prefix contract. The direct-stream diagnostic found two leading line feeds, so this is a retained client/harness formatting failure, not a proven model defect. Outcome: no promotion or valid strict120 speed claim for this profile.

2026-08-26 — Portable-QSA correctness baseline

The portable 262K TP=2 recipe established the same-day correctness baseline for the pinned RadixArk checkpoint. It passed the bounded service gates and was superseded later that day by the hardware-specific QSA-fast lane. It remains a slower historical control, not a transferable performance claim.

2026-08-26 — QSA-fast MTP3 qualification

The hash-gated SM120 QSA-fast path and its matched no-speculation control qualified the exact TP=2 runtime. NEXTN 3/1/4 improved matched decode 2.33x at 4K and 1.93x at 128K without a bounded quality regression. The full-reserve request passed at 253,703 actual prompt tokens with an 8,192-token output request inside the native 262,144-token window.

2026-08-26 — Multimodal and real-client acceptance

The hash-pinned 15-case corpus passed 30/30 direct. Routed and live repeats, router edge checks, and fresh OpenClaw, Hermes, and Pi turns established the dated text/image/OCR/video contract with fail-closed admission at four images or one video.

2026-08-30 — Retained rollback after the GLM evaluation began

The 2026-08-26 record had authorized this profile for llm.primary, vision.general, vision.ocr, and vision.video. When the GLM one-week evaluation began on 2026-08-30, Qwen3.8 Flash Next became the immediate retained text/image/OCR/video rollback. The label describes the August decision, not a current rollback assignment.

Immutable identity

Ignis 0.8 admission-only lane, October 24–25

Pinned Trellis revision 8d94db83cd8e58d6be31e796723e5e35b960638c, Ignis source 950ee0d1c2cf6178cd7fec29df3f17c3e127f987, and immutable image 2b8d56c75bbba758b504bc0ef7edea59bf45df398dd9a4cde1514edc81b63029 form a separate lane from the historical profiles below. Its TP1/131,072-token/eight-lane/MTP3 settings are configured values. R10 used 49 GiB RAM/no extra swap with a 16 GiB host reserve; the inherited 52 GiB prose note is an erratum. Exact pins and boundaries.

October Strata candidate, unqualified at this checkpoint

Unsloth GGUF revision 38bb39ee97821de2c9009abb7e93950eec396e66, UD-IQ4_XS; Strata 6f32ec070f23ced9f50e704d854d775da52591ab. The measured initial 32K/C1, 50 GiB containment/36 GiB resident-expert setup and separate instrumented image pins are in the follow-up identity. The earlier 52 GiB preparation proposal is retained separately; bounded startup is not final qualification.

September EXL3 scout

  • Checkpoint: turboderp/Qwen3.8-Flash-Next-exl3 at 55a732e0c4c3d4614bc42b68493bb930d9b02c0a, 4.05-bpw conversion.
  • Runtime image: sha256:00909876aea02112b75d775bfa50f6d1adac50a3a11918363e7b403d7340d9bc.
  • Identity evidence: configuration pins.

Checkpoint and runtime

  • Checkpoint: RadixArk/Qwen3.8-Flash-Next-NVFP4
  • Revision: 7b719225242aacd3dbd3f9407468c2ee9a9d2594
  • Served name: qwen38-flash-next-radixark-nvfp4-sglang-qsa-fast-tp2-262k-mtp3
  • SGLang image: sha256:59f06adce6f91401adf443bd168d45fdb2044d77671fd591c7c57a29d851cbae
  • Engine revision: d91c3682b0b429e4c70df63cd57f819588ce29b0

Quantization boundary

The checkpoint is a 180B-total multimodal MoE conversion. The tested quantization applies ModelOpt NVFP4 W4A4 to routed experts while attention, GDN, QSA, shared experts, vision, and MTP remain BF16; FP8 PLE tables are dequantized to BF16 at load. This dossier makes no lossless-quantization claim.

Tested hardware and topology

Qualified exclusive TP=2 lane

Two RTX PRO 6000 Blackwell Max-Q cards under WSL2, assigned exclusively to one TP=2 owner over PCIe without NVLink. Aggregate VRAM is 192 GB but is not unified memory. The retained MTP3 startup reported 6.275 GiB of KV cache per TP rank, 12.55 GiB aggregate, and a 516,032-token maximum server allocation; the matched no-speculation arm reported 415,744 tokens.

Single-card execution, another accelerator product, and co-resident serving were not tested for this retained configuration.

Engine, quantization, KV, context, and concurrency recipe

September EXL3 scout

The 262,144-token EXL3 profile is distinct from the August NVFP4 recipe below. Its native context artifact retains the executed controls. The managed recipe is operator-private; use the public reconstruction guide and the pinned evidence. Do not copy KV, speculation, or multimodal settings from the NVFP4 lane and label that a reproduction of the EXL3 scout.

Qualified QSA-fast MTP3 lane and matched control

The retained MTP3 recipe and matched no-speculation control pin the exact weights, image, SM120 QSA patch, TP=2, 262,144 context, BF16 KV, one running request, 0.80 static memory, backends, graph cap, parsers, transport, and no PLE CPU offload. Their only intended performance difference is SGLang NEXTN 3/1/4.

Exact-runtime compatibility controls

Startup hash-gates the exact QSA source and patched result from SGLang PR #36556 and fails unless the required FlashInfer decode symbol imports. The recipe also retains the qualified NCCL logits fallback, Triton attention/prefill, FlashInfer linear-attention decode, BF16 recurrent state, and batch-one CUDA graph capture. FP8 KV remains excluded. These are exact-revision SM120/WSL2 compatibility selections, not generic SGLang guidance. The portable 262K recipe is retained as the slower same-day correctness baseline.

Evidence by measurement class

R12 baseline and R13 admission refusal, 2026-10-25

  • Status: unexecuted candidate; host-policy admission refusal, no-promotion.
  • Configured, not measured: pinned Ignis 0.8/Trellis TP1 on one RTX PRO 6000, 131,072 context, eight profile lanes and MTP3; 49 GiB RAM/no extra swap plus a 16 GiB host reserve. These controls did not receive inference qualification.
  • Observed: the 65 GiB available-memory requirement was refused before creation; exact refusal-time available memory and shortfall are unknown. All 48 candidate jobs remained planned. No model traffic or router alias transition occurred.
  • Limits: no candidate grade, latency, throughput, peak memory or GPU-capacity result; an admission refusal is not an OOM or model-quality result. The prospective comparison retained all six failed R12 grades among 44 eligible baseline keys and excluded four cancelled unknowns, but no paired comparison ran.
  • Evidence: Authoritative summary, index, manifest, finding. Sanitized derivatives retain distinct original-private and public-byte hashes.

Ignis 0.8 startup and admission, 2026-10-24–25

R9 retains a pre-create host-envelope refusal. R10 retains loaded/ready startup evidence, a source-verified Anvil Serving identity-adapter defect and correct monitoring refusal. The coordinator's raw readiness/models responses were not durably retained; their provenance is explicit. Zero generation means no model quality, latency, throughput, energy or capacity result. Configured eight-lane concurrency is not measured capacity. Diagnostic finding and artifacts.

September EXL3 quality and strict capacity

  • Status: quality, functional, bounded context capacity; strict performance population ineligible.
  • Measured: MMLU-Pro 91/100, context 9/9 above 257K total reserved tokens, agentic 21/30, frozen SWE 4/5, image/OCR 12/12.
  • Limits: one quality pass, five SWE cases, and zero performance-eligible strict120 requests. The older NVFP4 video/client evidence does not transfer.
  • Evidence: dated finding and native artifacts.

Functional, capacity, and bounded quality

  • functional: direct coding, JSON, 128K retrieval, full-reserve request, tools 20/20, and exact routed identity/readiness/admission passed.
  • capacity: the MTP3 arm passed 253,703 actual prompt tokens with an 8,192-token output request inside the native 262,144-token window. The client contract remains 253,952 plus 8,192.
  • bounded quality: the thinking-disabled protocol-v3 suite passed intelligence 6/6, session 3/3, and tools 3/3.

Performance and context

  • performance: QSA-fast MTP3 measured 0.15 s TTFT, 0.36 s E2E, and 154.9 tok/s decode at 4K/c1; at 128K it measured 12.67 s TTFT, 9,664 prefill tok/s, and 134.1 decode tok/s. MTP3 improved matched decode 2.33x at 4K and 1.93x at 128K without a bounded quality regression.
  • context curve: 25/25 c1 requests completed at six targets from 4K through 254K. Median decode was 155.9 tok/s at 4K, 114.7 at 128K, and 112.9 at the 254K target with 245,000 actual prompt tokens. These 512-output-request rows remain separate from the 8,192-token full-reserve capacity proof.

Multimodal, routing, and clients

  • multimodal: the hash-pinned 15-case corpus passed 30/30 direct. Isolated router repeats scored 27/30 and 30/30; live repeats scored 29/30 and 28/30, or 57/60 strict. Every routed miss was a correct observation that omitted one literal rubric word. Router admission/SSE/tool/error edges passed 8/8.
  • live acceptance: fresh OpenClaw, Hermes, and Pi turns selected the Primary, returned the fresh marker without fallback, and Pi/OpenClaw reported the 262,144-token catalog contract. Their explicit vision references remain vision.general; fresh image-path acceptance is retained with the vision promotion.

See the QSA-fast MTP3 promotion record and vision-promotion record, plus their sanitized evidence bundles.

Decision and promotion state

September EXL3 challenger

The EXL3 scout remains no-promotion after the September 14 GLM selection.

Historical NVFP4 rollback

This was the retained video-capable rollback during the August GLM evaluation. Its historical promoted contract remains 253,952-plus-8,192, concurrency-one admission, four images, one video, and thinking disabled. The c2 queue diagnostic is not a c2 qualification. Restoring route metadata alone does not make the service live; a rollback requires the managed serve lifecycle plus fresh direct identity, route, sync, and real-client checks. A different media limit, KV dtype, worker/cache policy, or concurrent-media shape requires a fresh gate.

Transfer boundary

The portable QSA decoder remains a correctness-qualified historical control. The retained SM120 fast path and MTP3 result cannot transfer to another runtime, patch, KV dtype, offload policy, context, or speculation preset without matched functional, quality, capacity, and client gates.

Failures and gotchas

Runtime and compatibility

  • The pinned symmetric-memory logits path failed on SM120/WSL2; the qualified recipe uses the exact-revision NCCL fallback.
  • The pinned FA4 CuTe sparse decoder failed MLIR compilation; the qualified lane first used SGLang's device-agnostic QSA decoder, then moved to the exact hash-gated PR #36556 SM120 fast path.
  • Three exact Inferact/vLLM recipes were empirically disqualified by default NCCL initialization, V2-runner UVA, and a pre-KV V1 compile-autotune OOM. These failures do not prove universal checkpoint or 262K infeasibility.

Measurement and integration

  • The first long-tool generator undershot 100K measured tokens and was rerun with a calibrated prompt. That failed attempt is not counted as a pass.
  • Exact-context router admission, Responses bounded-thinking translation, Hermes catalog drift, and Pi provider seeding were fixed forward before the final client acceptance. These product/configuration fixes do not broaden the model qualification beyond the recorded c1 text/image/OCR/video contract.

Dated run history

  • 2026-10-25 — R12/R13 successor: Separate R13 candidate-only attempt refused before creation: 0/48 submitted. Incumbent restoration and minimal-thinking Pi accepted; thinking-off failure remains open; no candidate measurement.
  • 2026-10-24–25: Ignis 0.8 evaluation-platform admission diagnostics — R9 refused, R10 monitoring blocked after startup; candidate generation unmeasured, pilot acceptance incomplete.

  • 2026-10-13 — full qualification closure: QAD and EXL3 4.05 TP1 records retain distinct protocol and quality evidence; the Tabby parser correction has full-13 protocol and quality-12 coverage. The source-aligned top-k-20 control was cancelled before load/query, so no quality or speed improvement is claimed. Ignis R4 terminal SWE was 2/5 resolved, 2/5 graded, and three empty; serial C1 passed 6/6, the separate C8 preflight passed 6/13 with seven immediate 503 refusals, and 24 SWE engine-full retry notices all recovered. Pylint remained ungraded/empty; these retries do not explain the three empty patches. no-promotion.

  • 2026-10-05 — Strata/UD-IQ4_XS research and Swift baseline: 5090/96 GB initial feasibility; protocol/image checks passed, practical latency failed, instrumented diagnostics separate; no promotion.

Date Event Result
2026-10-05 Prospective primary replacement scout rejected at 3/5 against the 4/5 frozen coding floor; strict C4 0/4 performance-eligible; no-promotion for this evaluated profile only.
2026-09-22 RadixArk NVFP4 TP2, C4 configured admission, serial C1 replacement scout direct preflight 6/6; agentic 16/18 and 15/18 miss the campaign 18/18 gate; no-promotion
2026-09-13–14 EXL3 quality, context, agentic, SWE, and image scout no-promotion; strict120 performance ineligible
2026-08-26 Full multimodal corpus, context curve, and vision-route/client promotion then-current text/image/OCR/video Primary; direct 30/30; live 57/60 strict; edges 8/8; 25/25 context requests
2026-08-26 QSA-fast plus matched MTP3 qualification and fix-forward promotion then-current text Primary; 154.9 tok/s at 4K and 134.1 at 128K; direct/routed/client gates pass
2026-08-26 Portable-QSA 262K TP=2 qualification and initial promotion superseded same day; retained correctness and failure baseline