schuttdev/hipfire-qwopus-9b

Qwopus3.5-9B-v3 — hipfire quantized

Jackrong/Qwopus3.5-9B-v3 quantized for hipfire, a Rust-native inference engine for AMD RDNA GPUs.

Files

File	Quant	Size	Speed (5700 XT)	Notes
qwopus-9b.hf4	HF4 (4-bit)	4.5 GB	~41 tok/s	Faster, fits 8GB VRAM
qwopus-9b.hf6	HF6 (6-bit)	6.9 GB	~34 tok/s	Better quality, needs 8GB

Usage

# Install hipfire
curl -L https://raw.githubusercontent.com/Kaden-Schutt/hipfire/master/scripts/install.sh | bash

# Pull and run
hipfire pull qwopus:9b        # HF4 (default)
hipfire pull qwopus:9b-hf6    # HF6 (higher quality)
hipfire run qwopus:9b

About

Qwopus3.5-9B-v3 is a Qwen3.5-9B finetune by Jackrong. It uses the same DeltaNet hybrid architecture (linear attention + full attention layers) and ChatML chat template as the base Qwen3.5 model.

hipfire runs this model natively on AMD GPUs (RDNA 1-4) without ROCm runtime overhead. Kernels JIT-compile for the detected GPU architecture.

Hardware

Any AMD GPU with HIP SDK support:

• RDNA 1: RX 5500/5600/5700
• RDNA 2: RX 6600/6700/6800/6900
• RDNA 3: RX 7600/7800/7900
• RDNA 4: RX 9070
• APUs: Strix Halo, Strix Point

License

Same as upstream: Apache 2.0