Sign up

sakamakismile/Qwen3.6-35B-A3B-NVFP4

🤗 On Hugging Faceimage-text-to-textapache-2.044 GBsafetensors✓ Checksum-verifiedupdated 0d ago
Magnet

Qwen3.6-35B-A3B-NVFP4

NVFP4 quantized version of Qwen/Qwen3.6-35B-A3B — the latest Qwen MoE with 256 experts, 3B active parameters, and state-of-the-art coding/agentic performance.

67 GB → 21.9 GB. Single NVIDIA Blackwell GPU. 168 tok/s.

Why This Model

Qwen3.6-35B-A3B is the new king of the MoE class:

  • SWE-bench Verified: 73.4 — surpasses models 10x its active parameter count
  • Terminal-Bench 2.0: 51.5 — best-in-class agentic coding
  • QwenWebBench: 1397 ELO — real-world web task performance
  • 256 experts, 3B active — extreme sparsity = extreme speed
  • 262K-1M context — native 262K, extensible to 1 million tokens
  • Gated DeltaNet + Attention hybrid — next-gen architecture

At NVFP4, it runs at 168 tok/s on a single Blackwell GPU — faster than Gemma4 MoE (130 tok/s) with dramatically better benchmark scores.

Key Specs

| | |

|---|---|

| Base model | Qwen/Qwen3.6-35B-A3B |

| Architecture | Qwen3.5 MoE — 35B total, 3B active, 256 experts (8 routed + 1 shared) |

| Quantization | NVFP4 W4A4 (weights FP4, activations FP4, scales FP8) |

| Format | compressed-tensors (native vLLM support) |

| Tool | vllm-project/llm-compressor (main) |

| Calibration | 512 samples, ultrachat_200k, seq_len=2048, moe_calibrate_all_experts=True |

| Size | 21.9 GB |

| Max context | 262,144 tokens (native) |

| Requires | NVIDIA Blackwell GPU (SM 120), vLLM nightly (cu130) |

Quickstart

vLLM

vllm serve Lna-Lab/Qwen3.6-35B-A3B-NVFP4 \
    --max-model-len 32768 \
    --reasoning-parser qwen3 \
    --kv-cache-dtype fp8

With tool calling (agentic)

vllm serve Lna-Lab/Qwen3.6-35B-A3B-NVFP4 \
    --max-model-len 32768 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --kv-cache-dtype fp8

Docker

docker run --gpus '"device=0"' -p 8016:8016 \
    -v /path/to/model:/models/current:ro \
    --shm-size 16gb \
    vllm/vllm-openai:cu130-nightly \
    vllm serve /models/current --port 8016 --max-model-len 32768 \
    --reasoning-parser qwen3 --kv-cache-dtype fp8

Benchmark

Single NVIDIA RTX PRO 6000 Blackwell (96 GB VRAM).

| Test | Speed | Tokens | Result |

|------|-------|--------|--------|

| English (CAP theorem) | 161 tok/s | 256 | PASS |

| Code (async scheduler) | 162 tok/s | 512 | PASS |

| Math (Bayes' theorem) | 162 tok/s | 512 | PASS |

| Reasoning (architecture) | 163 tok/s | 512 | PASS |

| Container burst (x3) | 168 tok/s | 512 | PASS — stable |

Speed Comparison (NVFP4, single GPU)

| Model | Active Params | tok/s | Relative |

|-------|:------------:|:-----:|:--------:|

| Qwen3.6-35B MoE | 3B | 168 | 1.0x |

| Gemma4-26B MoE | 3.8B | 130 | 0.77x |

| Qwen3.5-27B Dense | 27B | 57 | 0.34x |

| Gemma4-31B Dense | 31B | 51 | 0.30x |

Quantization Details

Recipe

recipe = QuantizationModifier(
    targets="Linear",
    scheme="NVFP4",
    ignore=["lm_head", "re:.*visual.*", "re:.*mlp.gate$", "re:.*mlp.shared_expert_gate$"],
)

Calibration

  • Dataset: HuggingFaceH4/ultrachat_200k (train_sft split)
  • Samples: 512
  • Max sequence length: 2048
  • moe_calibrate_all_experts=True — ensures all 256 experts receive calibration data

Reproduction

from transformers import Qwen3_5MoeForConditionalGeneration, AutoProcessor, AutoTokenizer
from datasets import load_dataset
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier

MODEL_ID = "Qwen/Qwen3.6-35B-A3B"

model = Qwen3_5MoeForConditionalGeneration.from_pretrained(MODEL_ID, dtype="auto", trust_remote_code=True)
processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)

recipe = QuantizationModifier(
    targets="Linear", scheme="NVFP4",
    ignore=["lm_head", "re:.*visual.*", "re:.*mlp.gate$", "re:.*mlp.shared_expert_gate$"],
)

ds = load_dataset("HuggingFaceH4/ultrachat_200k", split="train_sft[:512]")
ds = ds.shuffle(seed=42)

def preprocess(example):
    return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)}
ds = ds.map(preprocess)

def tokenize(sample):
    return tokenizer(sample["text"], padding=False, max_length=2048,
                     truncation=True, add_special_tokens=False)
ds = ds.map(tokenize, remove_columns=ds.column_names)

oneshot(model=model, dataset=ds, recipe=recipe,
        max_seq_length=2048, num_calibration_samples=512,
        moe_calibrate_all_experts=True)

model.save_pretrained("Qwen3.6-35B-A3B-NVFP4", save_compressed=True)
processor.save_pretrained("Qwen3.6-35B-A3B-NVFP4")
tokenizer.save_pretrained("Qwen3.6-35B-A3B-NVFP4")

Environment

| Package | Version |

|---------|---------|

| torch | 2.11.0+cu130 |

| transformers | 5.5.4 |

| llmcompressor | 0.1.dev (main @ 3084520) |

| compressed-tensors | 0.15.1a20260414 |

| CUDA | 13.0 |

Requirements

  • GPU: NVIDIA Blackwell (SM 120)
  • VRAM: ~22 GB minimum (model only)
  • Software: vLLM nightly (cu130)

Notes

  • Multimodal (vision) preserved in BF16.
  • Gated DeltaNet layers are a hybrid attention+SSM architecture — unique to Qwen3.5/3.6.
  • NVFP4 is Blackwell-specific. Will not work on Ampere/Hopper.
  • Use --kv-cache-dtype fp8 for 2x KV capacity at no quality cost.

Credits