Qwen3.8 Flash Next¶

Current status and review date¶

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 2026-09-14; dossier reviewed 2026-09-17.

Latest finding · raw evidence · configuration history.

Review narrative¶

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 It passed the bounded service gates and was superseded later that day by the hardware-specific QSA-fast lane.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¶

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¶

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¶

Date	Event	Result
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