Three models. One principle: yours.

You get the actual model files. You can inspect them, run them, fine-tune them. There is no hosted version of our models that you have to go through us for. The weights live on your machine.

GGUF is a file format for AI models. It runs on Ollama, LM Studio, llama.cpp, and any compliant runtime. If you have a modern Mac, Windows, or Linux machine with the recommended RAM, you already have what you need. Download the file, point your runtime at it, done.

flash-1-mini is fine-tuned from an open, Apache-2.0-licensed Qwen base model — not Llama — and we attribute that base as the license requires (see the NOTICE file). That cuts both ways, honestly: the gains are real and measured against that exact base, and you still own what you download. Apache 2.0 is irrevocable — it lets you run, modify, and redistribute the model commercially, and nobody can pull the license out from under you. We fine-tune on Canadian infrastructure.

Yes. Outputs are yours. Use them in work product, client deliverables, internal tools, products you sell. The model license covers running the model. It does not claim ownership of what you generate.

Benchmarks publish with each model launch. flash-1-mini's are live now — see the model page for the full table, including the honest tradeoffs. We don't pre-announce numbers or cherry-pick; the regressions sit right next to the gains. Real numbers, on the actual model you can download.

You get a notification when a new version is available. You decide whether to download it. The version you already have continues to work forever — we cannot break or revoke a model you already downloaded.

Your models keep working. They are files on your hardware. We cannot revoke them, even by ceasing to exist. This is the entire point of open-weight ownership.

Yes. Once downloaded, the model runs entirely on your hardware. Airplane mode is fine. So is an air-gapped network.

GGUF runtimes work on macOS, Windows, and Linux. The September 2026 desktop app ships on macOS and Windows at launch, with Linux to follow.

No. We do not operate inference servers. We could not collect your prompts even if we wanted to. The desktop app, when it ships, sends no telemetry. See the privacy policy for full details.