Qwen3.8-27B — NVFP4 (ModelOpt)
NVFP4 quantization of Qwen/Qwen3.8-27B,
produced with NVIDIA TensorRT Model Optimizer 0.45
using NVIDIA's own recipe for the qwen3_5 family
(huggingface/qwen3_5/ptq/w4a16_nvfp4-fp8_attn-kv_fp8_cast) — the same
mixed-precision scheme NVIDIA used for its official Qwen3.6-27B NVFP4 release:
| Component | Precision |
|---|---|
MLP projections (gate/up/down) + lm_head | NVFP4 (W4A16, weight-only, block size 16) |
| Self-attention and linear-attention (GDN) projections | FP8 (E4M3, weights + activations) |
| KV cache | FP8 (constant amax) |
| Vision tower, MTP head, conv/gating layers | BF16 (unquantized) |
Checkpoint size is ~21 GB (vs. ~55 GB BF16).
Calibration
Post-training quantization with 1024 samples (max. 512 tokens each), 256 each
from four openly available datasets: cnn_dailymail,
Magpie-Align/Magpie-Pro-MT-300K-v0.1, nvidia/OpenCodeReasoning,
nvidia/OpenMathReasoning. Attention implementation during calibration: SDPA.
Serving with vLLM
vllm serve <this-repo> \
--quantization modelopt \
--max-model-len 262144 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder \
--enable-auto-tool-choice
NVFP4 runs natively on Blackwell GPUs; the checkpoint loads on earlier architectures via vLLM's ModelOpt support with reduced benefit.
Notes
- Community quantization by A2Genesis — not affiliated with or endorsed by the Qwen team or NVIDIA.
- The MTP (multi-token prediction) tensors are preserved in BF16, so speculative decoding remains available.
- License: Apache 2.0, inherited from the base model. All credit for the model itself belongs to the Qwen team.