---
datasets:
- nvidia/OpenCodeInstruct
base_model:
- cerebras/Qwen3-Coder-REAP-25B-A3B
---
# Qwen3-Coder-REAP-25B-A3B-nvfp4

**Format:** NVFP4 — weights & activations quantized to FP4 with dual scaling.  
**Base model:** `cerebras/Qwen3-Coder-REAP-25B-A3B`  
**How it was made:** One-shot calibration with LLM Compressor (NVFP4 recipe), long-seq calibration with nvidia/OpenCodeInstruct.  

> Notes: Keep `lm_head` in high precision; calibrate on long, domain-relevant sequences (256 samples at 4096 max length).

Check the original model card for information about this model.

# Running the model with VLLM in Docker

## ⚠️ Known vLLM Compatibility Issues

[This discussion](https://huggingface.co/Firworks/Qwen3-Coder-REAP-25B-A3B-nvfp4/discussions/1) contains a workaround to get this model running in VLLM until the below issues have been fixed/merged in VLLM.

This model currently does not work with vLLM due to CUTLASS FP4 kernel constraints - the 103-expert gate layer fails with `Expected n to be divisible by 32`. See [#24921](https://github.com/vllm-project/vllm/issues/24921), [#30934](https://github.com/vllm-project/vllm/issues/30934)

Additionally, Blackwell consumer/workstation GPUs (RTX 5090, RTX PRO 6000, DGX Spark) have a separate issue where FlashInfer FP4 GEMM doesn't support SM120/SM121. See [#31074](https://github.com/vllm-project/vllm/issues/31074), [#30163](https://github.com/vllm-project/vllm/issues/30163), [#23497](https://github.com/vllm-project/vllm/issues/23497)

```sh
sudo docker run --runtime nvidia --gpus all -p 8000:8000 --ipc=host vllm/vllm-openai:nightly --model Firworks/Qwen3-Coder-REAP-25B-A3B-nvfp4 --dtype auto --max-model-len 32768
```
This was tested on an RTX Pro 6000 Blackwell cloud instance.

If there are other models you're interested in seeing quantized to NVFP4 for use on the DGX Spark, or other modern Blackwell (or newer) cards let me know. I'm trying to make more NVFP4 models available to allow more people to try them out.