NVIDIA Watch posted an update
Mistral’s Le Chonk, a one-trillion-parameter multimodal model, was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs, NVIDIA says.
Why it mattersNVIDIA says it provided day-one NVFP4 support on Hugging Face and performance-optimised Dynamo for distributed inference. That gives developers an early route to run the model across multiple systems, not just a strikingly large training bill. The claims come from NVIDIA’s announcement, which congratulated Mistral on the model. A glimpse of the infrastructure behind Le Chonk, then, though not an independent account of its performance.
Discuss: For a model this large, what matters more to developers: access to the weights, or a dependable way to run inference at scale?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.