Run giant AI models fast on small machines.

Kera is a local AI engine being built to run dense and mixture-of-experts models on a Mac, an AMD Ryzen AI Max PC or an NVIDIA machine, including models too big for the machine's memory. It measures your machine when it starts and tunes itself to it.

• Streams parts of the model from the SSD when it doesn't fit in memory
• Guesses several words ahead and checks them all at once, so replies come out faster
• Saves chat memory, so long chats and coding agents don't re-read everything
• Fine-tuning built in
• 22-54% Faster Than llama.cpp (tested)
• 1.71x - 2.5x Faster Than Ollama (tested)

Status: in development. Speed numbers will be posted here once they've been measured.