Qwen3.6-27B NVFP4 for NInfer

This model card is the version-controlled source for neroued/Qwen3.6-27B-nvfp4-NInfer.

The repository contains the NVFP4 representation of Qwen3.6-27B, converted from the fixed packed weights in rdtand/Qwen3.6-27B-PrismaSCOUT-Blackwell-NVFP4-BF16-vllm to the native NInfer .ninfer artifact format. The artifact is intended only for NInfer; it is not a Transformers checkpoint, Safetensors distribution, or GGUF file.

The artifact uses the Qwen3.5 Dense architecture with mixed NVFP4/BF16 weights. Actual bindings and activation permissions select the native execution paths, including W4A4 Tensor Core prefill and A16 NVFP4 decode. Text, Vision, MTP, prefix reuse, CLI and serving use the common Engine route.

Artifact

The file contains Text, Vision, MTP, the optimized proposal head, and frontend resources. Text linears use the source repository's mixed NVFP4/BF16 allocation; vocabulary and MTP projections use Q8. Vision and speculative weights are loaded only when selected at startup.

Verify a downloaded file with:

printf '%s  %s\n' \
  '0448262d15df2ae4fda761540c110bc19e7c3b4f43c0938e4d50474429cda083' \
  'qwen3_6_27b_nvfp4.ninfer' | sha256sum --check

Requirements

• Current NInfer checkout, built from source;
• 64-bit Linux;
• NVIDIA GeForce RTX 5090 (sm_120a);
• Build dependencies from the build guide.

Already have the official v2 file? Upgrade it locally without downloading the weights again.

NInfer does not provide an install target or packaged binary. See the repository README for source-build dependencies.

Download and run a CLI example

hf download neroued/Qwen3.6-27B-nvfp4-NInfer \
  qwen3_6_27b_nvfp4.ninfer \
  --local-dir models

./build/rtx5090-release/apps/ninfer models/qwen3_6_27b_nvfp4.ninfer \
  --prompt "Explain prefill and decode in three sentences." \
  --max-context 32768 \
  --max-tokens 8192 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 3 \
  --proposal-head optimized

For images, videos, and structured chat history, see the CLI guide.

Start a local server

./build/rtx5090-release/apps/ninfer-serve models/qwen3_6_27b_nvfp4.ninfer \
  --host 127.0.0.1 \
  --port 8080 \
  --max-context 240000 \
  --context-capacity auto \
  --max-concurrency 2 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 3 \
  --proposal-head optimized \
  --preserve-thinking

Each request has a 240,000-token logical ceiling. KV histories and continuation state share the Device Context byte capacity selected at startup. The log reports the pool size and estimated token capacity. Requests allocate as they advance; resource pressure can reclaim retained prefixes or pause requests for later restoration/replay. See the startup reference for capacity, Host cache and deployment controls.

See the HTTP serving guide for the API surface and the resource scheduling reference for cache and admission semantics.

Supported use

The artifact supports:

• text generation in thinking and non-thinking modes;
• image, multi-image, video, and mixed multimodal messages;
• MTP speculative decoding with draft windows from one to five;
• BF16, INT8, FP8, NVFP4, and K8V4 KV cache;
• CUDA Graph decode and compatible-prefix reuse;
• startup-bounded small-scale concurrent serving with true batched decode;
• the NInfer CLI;
• OpenAI Responses Core, OpenAI Chat Completions, and Anthropic Messages serving.

Performance

The single-request serving measurements below were collected on an NVIDIA GeForce RTX 5090 with CUDA 13.1 compile/runtime and CUDA driver API 13.3. Requests were submitted serially to a persistent ninfer-serve process with CUDA Graph enabled, a 1,024-token prefill chunk, INT8 group-64 KV cache, and prefix reuse disabled. Each single-request value is the arithmetic mean ± sample standard deviation over five fixed seeds; server warm-up completes before the measured requests.

Concurrent MTP=3 decode saturation

The concurrent campaign uses one 293-token prompt followed by an 8,192-token generation per active request. Each concurrency point starts a fresh server with MTP3, INT8 group-64 KV, CUDA Graphs, a 16,384-token per-request context limit, and prefix reuse disabled. Aggregate throughput includes only complete one-second intervals whose actual decode batch remains equal to C. Each row is one sustained wave.

At C=8, the profile sustains 1,146.9 aggregate decode tok/s, or 5.67× its C=1 throughput.

Long-context baseline (MTP disabled)

At 7,680 prompt tokens this is 3.48× the prefill throughput of the published groupwise-int profile; at 260,096 prompt tokens it is 1.55×.

MTP=3 long-reasoning decode

Thinking was enabled and the output limit was 65,536 tokens.

MTP=3 cross-scenario decode

Each category contains three fixtures and five seeds per fixture (15 samples). Thinking was disabled and the output limit was 4,096 tokens.

See the full methodology and results, including metric definitions, comparison data, and the exact reproduction command.

Evaluation

The historical serving revision was b3d4d0f50b868711c62432bbd68e746217a2f49a. Current evaluation commands are documented in the evaluation guide. The results below retain their original profile.

The artifact was evaluated through NInfer's OpenAI-compatible serving route with thinking enabled, MTP=3, and a 262,144-token context limit. EvalScope 1.9.0 used 0-shot prompts, rule-based scoring, and one sample per problem with temperature 0.6, top-p 0.95, top-k 20, presence penalty 1.0, and seed 42. All 258 configured samples completed and were scored.

These are single-sample results under the stated NInfer evaluation profile, not pass@k scores.

Limits

• This artifact targets one RTX 5090, with one to eight active requests per Engine.
• KV and continuation state share the startup Context pool. Usable context depends on the selected KV type, enabled components, concurrency and available memory.
• NInfer returns generated tool calls to the client; it does not execute them.

Provenance

The artifact identity, summarized object inventory, and conversion provenance are published in artifact-manifest.json. The exact storage contract is maintained in the v3 container reference.

License

This NInfer artifact is distributed under the Apache License 2.0. The Qwen3.6-27B base repository and the NVFP4 source repository are also licensed under Apache-2.0. Users remain responsible for complying with the license and applicable laws.