What it is
A desktop app and inference engine for running open-weight language models locally, whether on Apple Silicon, an NVIDIA or AMD GPU, or a plain CPU. When a model is downloaded, it benchmarks kernel configurations on your chip and keeps the fastest one, and it adds speculative decoding, quantized KV cache, and shared prefix caches for concurrent sessions. Developers who run coding agents such as OpenCode, Codex, or Claude Code can connect them in one click or through an OpenAI-compatible API, and it can also run as a server for other machines on a network.