wtf is samosa-chat?

Samosa Chat is a desktop and terminal application that lets you run large AI chat models directly on your own computer, without sending anything to the cloud. It bundles three different models you can switch between: Qwen3.6 35B, a large model with image support, Bonsai 27B, a compact 1-bit reasoning model, and Ornith 1.0 9B, a smaller model tuned for coding and reasoning. Only one model loads into memory at a time, so switching models unloads the previous one first. The project is aimed at people who want a private chat assistant that works entirely offline once the models are downloaded. There is no cloud account to sign up for and no telemetry sent anywhere. The app can manage the whole lifecycle of a model: downloading it from Hugging Face, verifying it with checksums, checking that there is enough free disk space, and resuming a download if it gets interrupted partway through. Models can take up meaningful disk space, from about 4 gigabytes for the smallest option to 24 gigabytes for the largest, so a machine with enough storage and around 16 gigabytes of RAM is expected. Samosa automatically figures out how much conversation history a model can hold based on the hardware it is running on, and it can automatically summarize older parts of a conversation to keep chats going without running out of context space. Installation is done by building the project from source using a Makefile, on either macOS or Linux, and there is also a background server mode that exposes a local API compatible with the OpenAI chat format, so other tools could talk to the locally running model. Right now, the plain terminal chat command only supports the Qwen model, while the graphical app supports all three.