Repository navigation
Lightreason Studio for the pinned BF16 Qwen2.5-3B-Instruct model.
- Run Standard and Lightreason Exact with the target checkpoint alone. The optional Turbo runner uses separately supplied draft and adapter files.
- Build, edit, and compare live output. Experimental Patch edit applies a validated JSON replacement to the current page.
- Includes the retained GPU coordination research runner, benchmark instructions, and transparent Lightreason branding.
- Original Lightreason source is licensed under 0BSD. Qwen model downloads have separate research-only terms. The release includes no weights or trained adapter.
The RTX 4090 smoke test checked target identity, Standard streaming, an Exact token match, and one real JSON patch. See the README for setup and the benchmark notes for measurement scope.