Note: If you have a multi-GPU SM120 Blackwell system (RTX 50/Pro), try my vLLM fork to resolve P2P / TP=2 issues (Pending PR into upstream).
https://github.com/Gadflyii/vllm/tree/main
Qwen3-Coder-Next-NVFP4
NVFP4 quantized version of Qwen/Qwen3-Coder-Next (80B-A3B).
Model Details
| Property | Value |
|---|---|
| Base Model | Qwen/Qwen3-Coder-Next |
| Architecture | Qwen3NextForCausalLM (Hybrid DeltaNet + Attention + MoE) |
| Parameters | 80B total, 3B activated per token |
| Experts | 512 total, 10 activated + 1 shared |
| Layers | 48 |
| Context Length | 262,144 tokens (256K) |
| Quantization | NVFP4 (FP4 weights + FP4 activations) |
| Size | 45GB (down from ~149GB BF16, 70% reduction) |
| Format | compressed-tensors |
Quantization Details
Quantized using llmcompressor 0.9.0.1.
NUM_CALIBRATION_SAMPLES = 20
MAX_SEQUENCE_LENGTH = 2048
DATASET = "HuggingFaceH4/ultrachat_200k" (train_sft)
moe_calibrate_all_experts = True
# Layers kept in BF16
ignore = [
"lm_head",
"re:.*mlp.gate$", # MoE router gates
"re:.*mlp.shared_expert_gate$", # Shared expert gates
"re:.*linear_attn.*", # DeltaNet linear attention
]
Benchmark Results
MMLU-Pro
| Model | Accuracy | Delta |
|---|---|---|
| BF16 | 52.90% | - |
| NVFP4 | 51.27% | -1.63% |
Context Length Testing
Successfully tested up to 128K tokens with FP8 KV cache (Not enough VRAM to test any higher context).
Usage with vLLM
Requires vLLM with NVFP4 support (0.16.0+), Transformers 5.0.0+
#vllm Serving
vllm serve GadflyII/Qwen3-Coder-Next-NVFP4 \
--tensor-parallel-size 2 \
--max-model-len 131072 \
--kv-cache-dtype fp8
License
Apache 2.0 (same as base model)
Acknowledgments
- Qwen Team for the base model
- RedHatAI for the quantization approach reference
- vLLM Project for llmcompressor