The lab
OpenUnum is built and used here every day, on one mini PC. This is the proof: the numbers, the engine, the box and the record of what happened.
- ENGINE: 43.2 t/s median, 188 requests, 0 failed (as of 2026-11-10) more
- BOX: Bosgame M5 · Ryzen AI Max+ 395 · 128 GB more
- Bench: Our engine against the one it replaced: three wins, one tie, six losses after round 2, the method, and what we are trying next.
- Engine: unum-rt: what it is, the one chip it runs on, the story of how it was built, and how to ask for early access.
- System map: Every part of OpenUnum and the projects around it, checked against the repositories.
- Projects: Status and evidence for every project: what is live, what is experimental, what is private.
One Strix Halo box.
The lab runs on a Bosgame M5 mini PC: an AMD Ryzen AI Max+ 395 with 128 GB of memory shared between the processor and the graphics. That shared memory is the point: a model far too large for a consumer graphics card fits, with room left for the agents and the media tools.
- What it holds at once: about 68 GiB of model weights, two 131k-token slots, the agent runtime, file search and the site's live answers.
- What it cannot do: two heavy GPU jobs together. A video render and the main model take turns.
- What it taught us: memory locking matters more than any speed flag. The traps are written up with the full setup.
The full setup and the traps · Bosgame M5 product page