Qwen3.5-9B-autoround-NVFP4-linearattn-BF16

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64b93e6bd6c468ac7536607e/mj6xac74jHGLqymiovObc.png" alt="The Kaitchup -- AI on a Budget" style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;" /> <div style="display: flex; justify-content: center; gap: 0.5em; margin-bottom: 1em;"> <a href="https://kaitchup.substack.com/subscribe"><strong>Subscribe and Support</strong></a> </div> </div>

This is Qwen/Qwen3.5-9B quantized with AutoRound to NVFP4 with linear-attention layers kept in 16-bit. The model is compatible with vLLM (tested: v0.16.1rc1). Tested with an RTX Pro 6000. Currently under evaluation.

• Developed by: The Kaitchup

Instructions

uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
uv pip install git+https://github.com/huggingface/transformers.git

vllm serve [this model ID]  --max-model-len 262144 --reasoning-parser qwen3

Acknowledgments

Thank you Verda for providing the needed compute. I used their H200s. Verda is a European, AI-focused cloud and GPU infrastructure provider with sovereignty, sustainability, data privacy, and performance at its core. Check them out if interested.