Nanbeige4.3-33B-NVFP4
NVFP4 quantization of Nanbeige/Nanbeige4.3-33B @ c1093c501cc6, made with
llm-compressor and saved in compressed-tensors
format for vLLM.
- Weights and activations in FP4 (W4A4): E2M1 values in blocks of 16 with FP8 (E4M3) block scales and a per-tensor FP32 scale. Activation block scales are computed on the fly at inference.
- Kept in BF16: token embeddings,
lm_head, norms. - Size: 22.1 GB (BF16 original: 66.2 GB).
Quantized by WebBrain, an open-source (MIT) AI browser agent for Chrome and Firefox that reads pages, extracts data and automates tasks. It runs on this model locally through the vLLM endpoint below.
Quick quality check
| Model | GSM8K, first 100 test questions, greedy, non-thinking |
|---|---|
| BF16 original | 96.0% |
| NVFP4 (this repo) | 95.0% |
A small sanity check run right after quantization (vLLM 0.32.0, NVIDIA RTX PRO 6000 Blackwell Server Edition). It catches a broken quant; it is not a benchmark.
Usage
vllm serve webbrain-one/Nanbeige4.3-33B-NVFP4 \
--enable-auto-tool-choice \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder
- Native FP4 tensor-core path on Blackwell GPUs (SM100 / SM120). On older GPUs vLLM runs the weights as FP4 but skips activation quantization.
- On a single 32 GB card (e.g. RTX 5090), add
--kv-cache-dtype fp8 --max-model-len 32768and raise the context length based on the KV-cache size vLLM reports at startup. The full 512K context needs far more memory. - Chat template options (
enable_thinking,preserve_thinking,tool_call_format) and the recommended sampling settings are the same as the original: temperature 1.0 / top_p 0.95 for agentic and tool use, 0.6 / 0.95 for reasoning and chat. See the original card for details and benchmarks.
Recipe
| Scheme | NVFP4, all decoder Linear layers (364 = 52 layers x 7), lm_head ignored |
| Algorithm | round-to-nearest (QuantizationModifier) |
| Calibration | 256 samples, up to 2048 tokens (mean 552), rendered with the model's chat template: mlabonne/open-perfectblend (256) |
| Tools | llm-compressor 0.14.0.1, compressed-tensors 0.19.0, transformers 5.17.0, torch 2.11.0+cu128 |
| Hardware | NVIDIA RTX PRO 6000 Blackwell Server Edition, 3.1 min |
config.json is the original config plus quantization_config; tokenizer and generation files are
unchanged copies of the original.
License
Apache-2.0, same as the original model by the Nanbeige LLM Lab.