unum-rt, our inference engine
An inference engine written in Rust for one job: keep a large local model fast while several agents use it at once. It replaced a popular open-source engine in our production on 2026-11-10.
- One chip family: It runs only on gfx1151 today: the Radeon 8060S graphics inside AMD Strix Halo processors (Ryzen AI Max+ 395). No NVIDIA, no Apple, no other AMD graphics yet.
- Private for now: The source is not public. The benchmark method and every number, losses included, are.
- Where it wins: Decoding while a second request loads a prompt: 45.1 tokens/s against 12.4. It also wins on copy-heavy output and, since round 2, on decoding deep in a long conversation.
- Where it loses: Cold prompts, cached follow-ups and memory. Six of ten rows on the bench page go to the open-source engine; the prompt rows sit at 82 to 97% of it.
Ask for early access
If you have a Strix Halo machine and want to try the engine, tell us about it. Requests go to the founder, who answers each one by hand. There is no price and no promise of a date.
It currently runs only on gfx1151 (Strix Halo). Requests for other hardware are welcome as a signal, and will wait.
We store what you type here, and nothing else about you, to answer your request. It is deleted when you ask or after 180 days. It is never shared or used for anything else.
The request form needs JavaScript. Without it, write to collab@aiunum.com with your hardware and what you want to use the engine for.
The numbers · The story of the engine