Inference Snaps turn complex generative model deployment into a single command. They are built on the idea that AI runtimes should be portable, composable, and local-first. Whether you’re experimenting with Gemma 4, Qwen 3 Coder, or DeepSeek R1, you can instantly install a pre-optimized runtime and spin up an API endpoint locally. Each model snap includes all dependencies and is tuned for your silicon: CPU, GPU, or NPU.

The Inference Snaps Hugging Face organization is used to maintain the model weights going into Inference Snaps. The weights are typically split into small shards of 5GB or less.