| Family | Active Metal kernel package |
|---|---|
| GLM-5.3-Flash | glm/flash/v1/ |
| Nemotron 3.5 Lightning | nemotron/lightning/v1/ |
| Qwen3.8 dense | qwen/dense/v1/ |
| Qwen3.8 Flash Next | qwen/flash_next/v1/ |
| Gemma 4 | gemma/v1/ |
The version names the implementation, not the model release. Each family imports its active package;
incompatible versions can occupy separate directories. CUDA kernels live in the family's cuda/ package.
Snapshot fingerprints hash family code, active kernels and declared shared dependencies. Nemotron declares dense-Qwen attention as a shared dependency. Gemma declares dense-Qwen's lane matmul and Nemotron's row matvecs. The snapshot key also includes the prefill plan and the MLX, mlx-lm and TensorFold versions.
See the family guide for arithmetic, state and verification requirements. A multi-row kernel must match its own serial row before it can verify drafts.