Jan

Why it's ranked here

Apache 2.0 licensing, an official Microsoft Store listing, and installer builds for Windows, macOS, and Linux all coming out of one repository make this a genuinely shippable product rather than a research demo. The build system already accounts for multiple GPU vendors and ships mobile build targets alongside the desktop ones, and its test suite is split across the core library, the web front end, and the extensions rather than left as one catch-all script. That level of packaging discipline is uncommon among local-model desktop clients, and it is what puts this one ahead of thinner alternatives.

What's good

The engine build is genuinely hardware-aware: separate variants exist for plain CPU, for two different GPU vendors' acceleration APIs, and for two CUDA generations, so the app is not quietly assuming everyone owns the same graphics card. Packaging goes beyond a bare GitHub release to an official Microsoft Store listing and a Flathub build, which lowers the trust barrier for a non-technical person installing something that talks to a language model. A known Windows build failure caused by an overly long file path is also handled automatically, relocating the build output rather than leaving contributors to debug a cryptic compiler error themselves.

Tradeoffs

Building this from source on Windows is not a one-command affair: it needs a specific compiler front end, plus separate build and packaging utilities, all correctly set up and on the system path, and GPU acceleration needs yet another vendor toolkit layered on top. The shared library that the rest of the app imports from is intentionally thin, so most of the real behavior lives inside the extensions and the native shell rather than in one place a newcomer could read start to finish. That spreads the actual logic across a larger workspace than the download page lets on.

How to use it well

This fits someone who wants one app for switching between a private, offline model and a hosted one for heavier work, instead of juggling separate clients for each. Point other local tools at its bundled server instead of a cloud vendor's endpoint when you want a drop-in stand-in during development. If the goal is to change how the app talks to models rather than just use it, expect real time spent in the extension code and the native shell, not a quick settings tweak. It is a poor fit if what you actually need is a headless, scriptable server with no desktop interface at all.

Observed

License

Apache 2.0

Local API interface

Runs an OpenAI-compatible local API server for other applications to call

Agent interface

Supports Model Context Protocol integration for agentic capabilities

Build stack

Native shell is written in Rust and built with Tauri

Distribution channels

Distributed through an official Microsoft Store listing in addition to direct installers

Hardware acceleration

Engine builds are compiled in separate CPU, Vulkan, Metal, CUDA, and ROCm variants

Documentation site

Documentation is a separate Nextra-based static site, independent of the desktop app build