What it does
- Runs local inference across two AMD Radeon AI PRO R9700 GPUs, with 32 GiB of memory on each card.
- Switches between inference, still-image generation, and video generation as work needs them.
- Keeps a named serving profile and a rollback so experiments have a clear baseline.
- Records throughput with the workload, concurrency, sampling settings, and engine that produced it.
Krea and MiniMax H3 media from the current local lanes carry COMMUNITY conditions; each finished output still needs its full lineage checked.
What was measured
On 10 October, the MXFP4 model on radiance engine 1.3.0 averaged 635 aggregate output tokens per second with thinking enabled at 16 concurrent streams in the local three-round F69 harness. The same row averaged 79 tokens per second with one stream. Those figures describe that harness and configuration; they do not predict how quickly an arbitrary agent task will finish.
What I learned
Concurrency changes the question a benchmark answers. I want the record to tell me how much work was running, what settings it used, and whether the result was useful before I turn a fast number into a recommendation.
Status
11 October 2026: the current two-card rig is in use. The documented serving profile was promoted on 10 October; further host expansion is separate work. The infrastructure repository is private.