NVIDIA Runs 753B-Parameter Model Across Two Stations
A demonstration pairs two GB300 GPUs to serve the GLM-5.3 model through an OpenAI-compatible endpoint.
Updated on Nov. 13, 2026 in Artificial Intelligence

NVIDIA demonstrated its 753-billion-parameter GLM-5.3 NVFP4 model running across two connected DGX Stations. A related vLLM deployment reached 607.8 tokens per second with 16 concurrent requests on SPEED-Bench Coding.
Why it matters
The demonstration shows that a model of this size, which requires about 464 GB of GPU memory for its weights, can be served on a pair of GB300 systems linked by ConnectX-8 RDMA. Serving such a model from a standard API endpoint bears directly on how large mixture-of-experts models are deployed beyond large data centers.
The deployment split the model's tensors across two GB300 GPUs, each with 252 GB of HBM3e memory, using tensor parallelism and vLLM with prefix caching, an FP8 key-value cache and multi-token prediction. Mean request latency at 32 concurrent requests was 453.5 seconds, and the source does not report performance on other workloads.
The players
NVIDIA
NVIDIA is a American technology company known for designing graphics processors and AI computing hardware, including the GB300 GPUs used in this demonstration.
The details
Station A runs an OpenAI-compatible API endpoint while Station B operates as a worker system, with the two GB300 GPUs linked over ConnectX-8 RDMA. The model's mixture-of-experts design activates 40 billion of its 753 billion parameters for each generated token.
Timeline
November 13, 2026: TokenPost publishes its article on the NVIDIA GLM-5.3 demonstration.
The Tech Race
The demonstration extends NVIDIA's GB300 and ConnectX-8 hardware from single-system inference to a two-station setup serving a 753-billion-parameter mixture-of-experts model. It positions NVIDIA's DGX Station line as a route for running large open models outside hyperscale data centers.
Developers and organizations evaluating large AI models can see that a 753-billion-parameter model needs about 464 GB of GPU memory and can be served through a standard API endpoint on two linked stations. The source gives no pricing or availability for the DGX Station hardware.
The takeaway
A 753-billion-parameter model can be served on two linked workstation-class systems, but its throughput falls as concurrency rises, so deployment planning should weigh concurrency against latency.
Further reading
Explore more Artificial Intelligence coverage on Wisevoter.
Source note: This article includes information reported by TokenPost.


