CognitionKevin
Kevin-32B
Kevin-32B is a 32-billion-parameter fine-tune of Qwen’s QwQ-32B. It targets efficient CUDA kernels through multi-turn reinforcement learning.
Research release
How the model learns from execution
Each training turn receives earlier kernels, evaluation results and a short summary. Rewards depend on correctness and speed. The process compiles, executes and profiles candidate kernels.Cognition: Kevin-32B training
How to use it well
Practical guidance. Verify numerical correctness before comparing speed. Run the generated kernel on the intended GPU and input shapes; performance on one benchmark does not establish a general speed improvement.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | cognition-ai/Kevin-32B |