neroued/Qwen3.6-35B-A3B-NInfer

Qwen3.6-35B-A3B for NInfer

This model card is the version-controlled source for neroued/Qwen3.6-35B-A3B-NInfer.

The repository contains Qwen3.6-35B-A3B converted 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. Its optional DFlash companion weights come from z-lab/Qwen3.6-35B-A3B-DFlash.

Artifact

The file contains Text, Vision, MTP, DFlash, the optimized proposal head and frontend resources. Routed experts use Q4 gate/up and Q5/Q6 down weights; shared experts and mixer projections use Q8. Vision and speculative weights are loaded only when selected at startup.

Verify a downloaded file with:

printf '%s  %s\n' \
  '3e33297645dc33557751be1a3c407a74ed7c00f34909b5d4e8cfdce91b3dbe84' \
  'qwen3_6_35b_a3b.ninfer' | sha256sum --check

Requirements

• —NInfer revision `98dada0` or later, built from source;
• —64-bit Linux;
• —NVIDIA GeForce RTX 5090 (sm_120a);
• —CUDA Toolkit 13.1 or newer.

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/Qwen3.6-35B-A3B-NInfer \
  qwen3_6_35b_a3b.ninfer \
  --local-dir models

./build/apps/ninfer models/qwen3_6_35b_a3b.ninfer \
  --prompt "Explain prefill and decode in three sentences." \
  --max-context 32768 \
  --max-new 8192 \
  --kv-dtype fp8 \
  --spec mtp --draft-tokens 3 \
  --lm-head-draft

For DFlash, the measured block-8 configuration uses seven draft tokens:

./build/apps/ninfer models/qwen3_6_35b_a3b.ninfer \
  --prompt "Explain prefill and decode in three sentences." \
  --max-context 32768 \
  --max-new 8192 \
  --kv-dtype fp8 \
  --spec dflash --draft-tokens 7 \
  --lm-head-draft

--draft-tokens accepts 1..5 for MTP and 1..15 for DFlash. The DFlash value 7 is the measured block-length-eight profile; 15 uses the companion's full native 16-position block. MTP and DFlash are mutually exclusive backend selections. DFlash may be combined with --vision for image or video prompts; it accelerates generated-text decode, not Vision encode or target prefill.

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

Start a local server

./build/apps/ninfer-serve models/qwen3_6_35b_a3b.ninfer \
  --host 127.0.0.1 \
  --port 8080 \
  --max-context 240000 \
  --kv-capacity 240000 \
  --max-concurrency 2 \
  --kv-dtype fp8 \
  --device-state-slots 2 \
  --host-state-slots 8 \
  --host-kv-mib 8192 \
  --spec mtp --draft-tokens 3 \
  --lm-head-draft \
  --preserve-thinking

Each request has a 240,000-token logical ceiling. The shared 240,000-token Device KV pool admits two active requests when their combined completion reservations fit; either request may use the full pool while running alone. Two extra Device checkpoint slots, eight pinned Host State slots, and 8 GiB of pinned Host KV retain reusable continuations under resource pressure.

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;
• —DFlash speculative decoding for Text and image/video Vision prompts with draft windows from one to fifteen;
• —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

Measured on 2026-09-07 at NInfer revision 487f89773f07cb18a2fb841fe0971ec9634d409b through the public HTTP serving route on one RTX 5090. The Release build uses CUDA 13.1 compile/runtime and CUDA driver API 13.3, INT8 group-64 KV, CUDA Graphs, 1,024-token prefill chunks, and disabled prefix reuse. Each measurement point starts a fresh server and excludes startup warmup.

Long-context baseline

No speculative backend; context ceiling 262,144 tokens; five fixed seeds per prompt length. Values are arithmetic mean ± sample standard deviation.

MTP3 single-request decode

Three draft tokens, optimized proposal head, stochastic sampling. The C=1 corpus point supplies five samples per reasoning fixture and fifteen per scenario category. Reasoning enables thinking with a 65,536-token output budget; other scenarios disable thinking with a 4,096-token budget.

MTP3 corpus makespan

Each C runs the same 75-request corpus. Rates use the full makespan, including prefill, workload transitions, and drain. The per-request context ceiling is 262,144 tokens and KV capacity is automatic.

DFlash K=7 single-request decode

Seven draft tokens, optimized proposal head, and the same corpus and context ceiling. Values below use stochastic sampling and per-request decode-phase timings.

MTP3 decode saturation

One 8,192-token generation per active request, a 16,384-token context ceiling, and one wave per C. Steady rates include only complete intervals with a full decode batch; acceptance covers the full wave.

All 485 formal requests across the 11 measurement points completed without request, CUDA, or out-of-memory failures. No obvious short-cycle repetition was found in the 225 C=1 speculative responses. Output-limit samples remain in the reported results.

See the complete performance report for DFlash greedy results, corpus token totals, termination counts, resource settings, and commands.

Evaluation

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 configured samples completed and were scored.

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

Limits

• —NInfer executes on one RTX 5090 and one CUDA device, with a startup-fixed capacity of 1–8 active requests per Engine.
• —It does not provide large-scale or preemptive continuous batching, priority/QoS scheduling, multi-GPU execution, CPU/GPU offload, or distributed serving.
• —Context allocation is subject to GPU memory and the selected KV-cache type.
• —NInfer does not execute generated tool calls.

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 source Qwen3.6-35B-A3B repository is also licensed under Apache-2.0. Users remain responsible for complying with the license and applicable laws.