MASHBOX Labs

MASHBOX Labs makes open models run faster and cheaper on TPUs and accelerated hardware, as well as exploring the future of the intersection of AI and gaming.

mashbox.gg • GitHub • Docs • Supported models • Use with coding agents • X • hello@mashbox.gg

We make open models run faster and cheaper on Google Cloud TPU, and we build AI for games.

The models here are our own builds. <Model>-TPU repos in the "TPU models" collection are open models we port to Google Cloud TPU, and they run there with the mashbox package; some of them also use our own TPU kernels. <Model>-TPU repos in the "Weights for TPU" collection hold files made for Google Cloud TPU that mashbox run does not start; each card shows how to use them. <Model>-CoreML or <Model>-ane repos in the "Apple Neural Engine models" collection hold files made for the Apple Neural Engine that mashbox run does not start; each card shows how to use them. We also trained a rigging model of our own. Every port's card names the original model and its makers and keeps the original licence.

Our engine, mashbox, runs these builds and also runs GGUF files from the community as they are. Families covered so far: Qwen, Llama, Gemma, Granite, Falcon and EXAONE. Every file, and how far it has been checked, is on Supported models.

On a Google Cloud TPU VM (the setup guide gets you there from a new Google account):

python3 -m pip install uv && python3 -m uv tool install --python 3.12 "mashbox[tpu]"
~/.local/bin/mashbox run qwen3-8-27b

Works with Claude Code, Codex and Pi. The mashbox server speaks the Anthropic Messages, OpenAI Responses and OpenAI chat completions APIs, so coding agents can use the LLMs here as their model: Use with coding agents.