Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
Magnitude is an open-source local inference engine built around a narrower problem than most local-LLM runners try to solve: figuring out, on the specific machine in front of you, which models will actually run well and then getting them running without manual GGUF hunting or quantization guesswork. Instead of asking a user to already know their hardware's limits, the magnitude setup CLI command profiles the machine -- chip, memory, and bandwidth -- and ranks candidate models by estimated speed, accuracy, intelligence, and memory footprint so the developer picks from a shortlist tuned to what their hardware can sustain, rather than trial-and-error downloading multi-gigabyte weights that turn out to be too slow or too large. Once a model is selected, Magnitude downloads it, applies hardware-specific tuning such as speculative decoding, and wires the result into whichever agent harness the developer is already using, rather than requiring a separate integration step per tool: the README lists Pi, OpenCode, Hermes, OpenClaw, Codex, Claude Code, Oh My Pi, and Cline as supported harnesses, with setup able to be delegated to the coding agent itself by pasting a short onboarding prompt that has the agent install the CLI, run magnitude docs onboarding, and switch itself over to the chosen local model. The engine runs models on demand -- loading a model into memory only when an agent actually calls it, and unloading it when the agent goes idle or system memory pressure rises -- so a developer isn't left running a multi-gigabyte model in the background between requests. It targets macOS, Linux, and Windows (via WSL) across Apple Silicon, NVIDIA GPUs, AMD GPUs, and CPU-only machines, and the project frames the resulting setup as fully private: once weights are downloaded, prompts, files, and inference stay on-device with no further network dependency, no token costs, and no API keys. The codebase is primarily TypeScript with a Rust core for the inference/runtime layer, distributed under the Apache License 2.0, and ships as an installable npm CLI package (@magnitudedev/cli) alongside the open GitHub repository.