Qwen3.8 Flash Next NVFP4 262K qualification and promotion

• Date: 2026-08-26
• Decision: human-authorized reference promotion
• Measured hardware: 2× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, exclusive TP=2 over PCIe without NVLink under WSL2
• Evidence: sanitized summary

Outcome

RadixArk/Qwen3.8-Flash-Next-NVFP4 revision 7b719225242aacd3dbd3f9407468c2ee9a9d2594 is the human-authorized text Primary reference at 262,144 served tokens, concurrency one, and an 8,192-token client output reserve. The endpoint is served as qwen38-flash-next-radixark-nvfp4-tp2-262k through digest-pinned SGLang sha256:59f06adce6f91401adf443bd168d45fdb2044d77671fd591c7c57a29d851cbae. The engine revision is d91c3682b0b429e4c70df63cd57f819588ce29b0.

Authenticated routed coding, structured JSON, 253,325-prompt-token retrieval, tools 20/20, streaming tools, tool-result continuation, and the Responses API passed. Real OpenClaw, Hermes, and Pi turns selected llm.primary, completed through the local route, and did not fall back. All three clients now advertise the same 262,144-token context and 8,192-token output cap from authenticated router metadata; OpenClaw and Pi retained their compaction settings, and Hermes retained its selected Anvil provider. DeepSeek is no longer the reference text Primary. Active host assignments and raw client evidence remain private operator state.

Exact recipe

The public 262K recipe and conservative 32K first-load recipe pin the model revision and container digest. The promoted lane uses:

• two equal 96 GB cards as one exclusive TP=2 owner;
• NVIDIA ModelOpt NVFP4 W4A4 routed experts, with BF16 attention, GDN, QSA, shared experts, vision, and MTP tensors;
• 262,144 context, 253,952 prompt budget plus 8,192 output reserve, one running request, page size 64, 4,096-token chunked prefill, and 0.80 static memory;
• no MTP/speculative decode, no automatic truncation, and thinking disabled for the qualified Chat Completions contract; and
• a fail-closed NCCL logits fallback plus SGLang's device-agnostic QSA sparse decoder for this SM120/WSL2 compatibility lane.

The engine reported a 416,064-token KV pool, or 1.587 full configured windows. Router admission remains one full-window request. The 192 GB figure is aggregate sharded VRAM, not unified memory.

Qualification results

Functional and context

Gate	Result
Direct 262K client contract	exact retrieval at 253,325 prompt tokens in 17.606 s with 8,192 output-token budget and zero reasoning tokens
Direct long tool use	valid tool call after 105,756 measured prompt tokens in 4.664 s
Routed Primary	coding, JSON, 253,325-token retrieval in 28.38 s, tools 20/20, streaming tools, and tool-result continuation all passed
Responses API	completed with exact READY; default request emitted no reasoning
Client catalog	Hermes, Pi, and OpenClaw each declare 262,144 context tokens and 8,192 max output tokens for llm.primary
Real clients	OpenClaw, Hermes, and Pi passed exact route/model acceptance without fallback
Native boundary	262,137 prompt tokens plus one output accepted; a forced-output probe reached 262,142 visible sequence tokens; 262,138 prompt tokens was the first rejected size

The client contract deliberately reserves output space instead of advertising the engine's last prompt position as usable client context.

Repeated behavior and capacity

The thinking-disabled protocol-v3 gate passed intelligence 6/6, session recall 3/3, and typed tools 3/3. This is a bounded behavioral gate, not a general intelligence ranking or a claim that NVFP4 is lossless.

Workload	Completion	TTFT	E2E	Decode
4K, c1, five requests	5/5	214.612 ms p50	2.627 s p50	12.801 tok/s p50
4K, c2 queue diagnostic, four requests	4/4	2.822 s p50	5.290 s p50	12.445 tok/s p50
128K, c1, one request	1/1	11.961 s	13.795 s	12.544 tok/s

The c2 diagnostic completed but queued behind the one-running-request scheduler. It supports the concurrency-one admission decision; it is not a qualified c2 service contract. Decode is materially constrained by the portable sparse-attention compatibility path and no speculative decoder was enabled.

Fix-forward record

The first boots and promotion probes exposed real failures; none were hidden or converted into a pass:

1. SM120/WSL2 symmetric-memory logits initialization failed. The managed recipe now selects the existing NCCL fallback through an exact, fail-closed source anchor.
2. The pinned FA4 CuTe sparse decoder failed MLIR compilation on SM120. The qualified recipe selects SGLang's device-agnostic QSA sparse decoder while leaving the weights and prefill path unchanged.
3. The original long-tool generator did not produce 100K measured tokens. A calibrated 125K-target rerun passed at 105,756 measured tokens.
4. Router admission rejected an exact upstream context contract and the Responses adapter rejected bounded thinking control. Both product defects were fixed with regression coverage before the final routed rerun.
5. Hermes had credential drift and a stale 393,216-token model override, while Pi lacked a usable Anvil provider reference. The managed client synchronizer now reconciles Hermes, Pi, and OpenClaw from one authenticated router snapshot, uses environment-key references, preserves client compaction and unrelated settings, and is idempotent. Fresh real Hermes, Pi, and OpenClaw requests then passed.

A separate failed final rerun using a stale operator-shell credential is retained privately as authentication evidence. The successful public claim is grounded only in the authenticated rerun and sanitized summary.

Boundaries

• Multimodal weights are present, but this promotion qualified the text Primary contract; it does not promote a new image, OCR, or video route.
• The bounded repeated suite passed, but no broad model-quality benchmark or same-prompt BF16/FP8/NVFP4 comparison was performed.
• The recipe is pinned to this checkpoint, engine revision, image digest, SM120/WSL2 topology, TP size, and compatibility patches. Do not transfer the result to another Qwen revision, runtime, GPU, quantization, or speculation setting.
• Raw operator artifacts, addresses, credentials, GPU UUIDs, and client configuration remain outside the public repository. The linked summary is the bounded public evidence for the claims above.