35B parameter model. Running locally on a 12GB RTX 5070 laptop. 55 tokens/second.
For comparison, that's roughly in range of what you'd get from Claude or ChatGPT over API.
Powered by FreeToken, just published on GitHub.
Qwen3.6-35b-A3B NVFP4 is what I'm using but they added support for Qwen3.8 too.
One command hooks it into OpenCode: "ft launch opencode." Any other agent works too — sky's the limit.
No subscription. No API key. Just your GPU.
Feel free to reach out for any help with the setup.
Top comments (0)