metadata
base_model:
- Qwen/QwQ-32B
Kevin-32B
Kevin (K(ernel D)evin) is a 32B parameter model finetuned to write efficient CUDA kernels.
We use KernelBench as our benchmark, and train the model through multi-turn reinforcement learning.
For the details, see our blogpost at https://cognition.ai/blog/kevin-32b
