Qwen3.5 122B NVFP4 Primary qualification¶

Observed: 2026-07-28 Host: Primary Node, Windows/WSL2, one NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (96 GB, sm_120) Decision: qualified for a human-gated Primary promotion, with Laguna S 2.1 retained as the immediate managed rollback.

Exact identity and source review¶

The tested checkpoint is nvidia/Qwen3.5-122B-A10B-NVFP4 at revision 98915d837c4e7c87ac8296d02e89de19b3207e6d. The Hugging Face API reported that revision as the repository head on 2026-07-28; the repository was last modified 2026-06-02 and the model card identifies a 2026-06-01 release. This review found no newer official NVIDIA checkpoint.

The checkpoint config.json declares max_position_embeddings=262144. Therefore the exact native limit is 262,144 tokens, commonly rounded to 262K or approximately 264K. The local cache contained the complete 77.8 GiB pinned snapshot before startup.

NVIDIA's card provides an NGC vLLM 26.04 reference command with ModelOpt FP4, FP8 KV cache, tensor parallelism one, and Qwen reasoning/tool parsers. The qualified local recipe uses the newer pinned nvcr.io/nvidia/vllm:26.06-py3 image and BF16 KV cache. A single-RTX-PRO-6000 community recipe and a Hugging Face operator discussion independently report the full 262,144-token shape.

MTP was deliberately not enabled. It is absent from NVIDIA's reference command, and current reports include zero draft-token acceptance and checkpoint-specific output-quality concerns. An official-checkpoint discussion also reports occasional extended or unproductive reasoning. These community reports are recipe and risk priors, not local benchmark results.

Qualified serve¶

Setting	Qualified value
Image	`nvcr.io/nvidia/vllm:26.06-py3`
Runtime observed	vLLM `0.22.1+7b9cb5b7.dev`
Quantization / KV	`modelopt_fp4` / BF16
Served name	`qwen35-122b-a10b-nvfp4`
Context / admission	262,144 tokens / one sequence
Batch-token cap	8,192
GPU memory utilization	0.94
Tools / reasoning	`qwen3_coder` / `qwen3`
Multimodal policy	one image per prompt; video disabled
Production thinking policy	enabled by default; per-request disable supported

The final multimodal startup completed in approximately 203 seconds. vLLM reported 73.22 GiB of model memory, 13.85 GiB available for KV cache, 571,950 GPU KV-cache tokens, and 2.18-times maximum concurrency at the full 262,144-token window. Host telemetry observed approximately 88.5 GiB of VRAM in use.

Independent gates¶

Thinking-disabled preflight passed smoke, structured JSON, approximately 240K retrieval, and 10/10 tool calls.
Thinking-enabled preflight passed smoke, JSON, approximately 128K retrieval, and 3/3 tools while retaining visible reasoning evidence.
With the vision tower enabled, general image understanding and verbatim OCR passed with thinking disabled and enabled. Default-mode smoke and JSON also produced the required reasoning evidence.
After managed promotion, the direct Primary endpoint repeated image and OCR successfully with default thinking and with thinking explicitly disabled. The authenticated llm.primary route also passed both image checks. Its response adapter omits the upstream reasoning field, so direct-serve evidence is the authoritative check that default thinking executed.
The repeated protocol-v3 suite passed chat, context at 32K, 128K, and 240K, tools, session recall, unified-diff formatting, and timeout triage at the required 100% pass rate.
A separate near-ceiling 240K request completed with 69.58 seconds TTFT and 70.41 seconds end-to-end latency.
Short-output capacity completed 10/10 at concurrency one. The warm repeat measured 0.15/0.26 seconds TTFT p50/p95 and 65 output tokens/second aggregate.

Raw artifacts:

Caveats and rollback¶

The runtime labels the ModelOpt FP4 checkpoint format and Mamba align-mode prefix caching as experimental. Image input is deliberately bounded to one item, video is disabled, and thinking requires completion headroom. Near-ceiling prefill is functional but slow. The quality suite is a deterministic serving-contract gate, not a broad claim that Qwen is generally more intelligent than Laguna.

The promotion plan changes only the Primary service and llm.primary route. poolside/Laguna-S-2.1-NVFP4 remains pinned at revision 07614121b31898586430f189d27a25a0be310843 as primary-laguna-rollback; the managed transaction restores it if the Qwen startup, 240K gate, router reload, or routed identity check fails.