Qwen3-Coder-Next-NVFP4-GB10

gdubicki/Qwen3-Coder-Next-NVFP4-GB10

Public mirror of saricles/Qwen3-Coder-Next-NVFP4-GB10.

Weights are byte-identical to the upstream quant (config.json and model.safetensors.index.json SHA-256 verified). This mirror exists to provide a pinned, stable, ungated reference for the qwen3-coder-next deployment project on DGX Spark (GB10). Use the upstream repo if you want to track author updates.

Credits

• Base model: Qwen/Qwen3-Coder-Next by Alibaba / Qwen team (Apache-2.0)
• NVFP4 quantization: saricles using LLM Compressor with LLMCOMPRESSOR_MOE_CALIBRATE_ALL_EXPERTS=1 (all 512 experts calibrated)
• Calibration data: HuggingFaceH4/ultrachat_200k (64 samples × 2048 tok)
• License: Apache-2.0 (inherited from base model; redistribution permitted)

Model details

• Architecture: qwen3_next — Hybrid DeltaNet linear attention + full attention + latent MoE
• Layers: 48 total (36 DeltaNet linear-attention, 12 full-attention)
• Parameters: 79.7B total, ~3B active per token (512 experts, 10 active + 1 shared)
• Quantization: NVFP4 via compressed-tensors; lm_head, embed_tokens, linear_attn layers, mlp.gate, mlp.shared_expert_gate kept in BF16
• Size on disk: 45.9 GB (70% reduction from ~149 GB BF16)
• KV cache: FP8 (12 full-attention layers only; DeltaNet has no KV cache)
• Max context: 262,144 tokens

Verified performance

Measured on DGX Spark (GB10 Blackwell, SM12.1, 128 GB unified LPDDR5X) with vLLM (Marlin NVFP4, FP8 KV cache):

Same decode throughput as Nemotron-3-Nano-30B-A3B-NVFP4 (both 3B active), but with a native 262K context window.

Usage

docker run --rm --runtime=nvidia --gpus all \
  -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -e VLLM_NVFP4_GEMM_BACKEND=marlin \
  -e VLLM_TEST_FORCE_FP8_MARLIN=1 \
  -e VLLM_USE_FLASHINFER_MOE_FP4=0 \
  -e VLLM_MARLIN_USE_ATOMIC_ADD=1 \
  vllm/vllm-openai:cu130-nightly \
  gdubicki/Qwen3-Coder-Next-NVFP4-GB10 \
  --dtype auto \
  --kv-cache-dtype fp8 \
  --gpu-memory-utilization 0.90 \
  --max-model-len 262144 \
  --attention-backend flashinfer \
  --tool-call-parser qwen3_coder \
  --enable-chunked-prefill \
  --enable-prefix-caching

Full deployment scripts: https://github.com/grzegorzdubicki-ai/qwen3-coder-next