607.8 tokens per second… NVIDIA GLM-5.3 distributed across two devices
Original headline초당 607.8토큰…엔비디아 GLM-5.3 두 장비 분산 구동
Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.
NVIDIA reported that a distributed deployment of Z.ai's GLM-5.3 across two GB300 DGX Station systems sustained 607.8 tokens per second across 16 concurrent requests, with one station exposing an OpenAI-compatible API and the second acting as a worker node. GLM-5.3 is a 753-billion-parameter mixture-of-experts model that activates 40 billion parameters per token, and the test used an NVFP4-quantized checkpoint of roughly 464GB—larger than the 252GB of HBM3e on a single GB300 GPU, which is why the model had to be split across two systems. The result matters because it shows a frontier-scale MoE model can be served at interactive throughput on workstation-class hardware rather than a full data-center rack, broadening who can host such models. What to watch next is whether NVIDIA publishes latency and cost-per-token figures alongside throughput, and whether similar multi-station splits scale to larger models and higher concurrency.
Muhabbit's AI wrote this summary in its own words from TokenPost's report and translated it from Korean. It can get details wrong, so check the source.
Named in this story
NVIDIA (NVDA), GLM-5.3, GB300 DGX Station, Z.ai
A linked name opens that asset's sentiment page.