The model is based on Qwen2.5-Coder-7b it seems. I currently run some quantized variant of Qwen2.5-Coder-7b locally with llama.cpp and it fits nicely in the 8GB

The model is based on Qwen2.5-Coder-7b it seems. I currently run some quantized variant of Qwen2.5-Coder-7b locally with llama.cpp and it fits nicely in the 8GB VRAM of my Radeon 7600 (with excellent performance BTW), so it looks like it should be perfectly possible. I would also only use Zeta locally.