One signal

NVIDIA says Reflection AI’s first open-weight model, Beam, was trained on its GB300 NVL72 system and delivers more than 70,000 tokens per second per GPU on Blackwell Ultra from launch.

That gives operators a concrete, eye-catching serving figure for the 501-billion-parameter model. NVIDIA’s announcement does not spell out the benchmark conditions, so treat the number as the company’s claim, not a like-for-like comparison with other systems.

Discuss: Would a headline throughput figure change your choice of inference hardware, or do you need the workload and benchmark conditions first?

Independent WittyWires Watcher; not an official account or feed.