unum-rt
unum-rt is the engine that runs the lab's main model. It is written in Rust with GPU kernels in HIP, it loads one model family on one chip (gfx1151, the Radeon 8060S graphics inside AMD Strix Halo processors), and it was built to keep that model fast while two OpenUnum agents use it at the same time. It took over production from a popular open-source engine on 2026-11-10. The source is private. The benchmark, with the rows it loses, is public.
- Decoding while a second request loads a 16k prompt (round 2, 2026-11-11): unum-rt 45.1 tokens/s vs a popular open-source engine, 12.4 tokens/s (evidence)
- Decoding 256 tokens at 28k depth (round 2, 2026-11-11): unum-rt 61.1 tokens/s vs open-source engine, 53.4 tokens/s (evidence)
- Cold prompt, about 29.5k tokens: the row we lose by most (round 2, 2026-11-11): unum-rt 1,001 to 1,021 tokens/s vs open-source engine, 1,209 to 1,238 tokens/s (evidence)
- Graphics memory in use (round 2, 2026-11-11): unum-rt 81.4 GiB vs open-source engine, about 75 GiB (round 1 figure) (evidence)