BAAI's PhysisServe reports 4.1 ms latency for 33B MiniMax H3 world model serving
Article timeline
The articles in this story. Times are ET.
BAAI@BAAIBeijingOfficialPhysisServe scales world model serving through pipelining and schedulingAIBAAI introduces PhysisServe, a serving system that uses stage-wise pipelining, time-division multiplexing, and virtual rank scheduling to co-scale world model size and serving capacity. The post links to a technical report with full details and references.
- BAAI@BAAIBeijingOfficialPhysisServe reports 4.1 ms latency for 33B MiniMax H3 across five backbones
AIBAAI says PhysisServe, evaluated across five ACWM and WAM backbones with full weights and no pruning, compression, or quantization, reaches 4.1 ms single-stream output latency for the 33B MiniMax H3 with one GPU per stage. The post says this equals about 245 chunks per second, or roughly 1,000 FPS as an ACWM and up to 3,900 Hz as a WAM. The evaluated backbones are Wan2.1 1.3B, Wan2.2 TI2V 5B, Wan2.1 14B, LTX-2.5 22B, and MiniMax H3 33B.
- BAAI@BAAIBeijingOfficialPhysisServe maps virtual ranks onto GPUs for concurrent interactive streams
AIPhysisServe encapsulates each stream's compute tasks and runtime state into virtual ranks that are dynamically mapped onto physical GPUs. The scheduler places a new stream based on remaining GPU capacity and admits it only if all of its stages fit, without migrating running streams. This aims to keep service stable while balancing cross-GPU load and using fragmented idle capacity.
