Mistral AI and CoreWeave Showcase Their Partnership at NVIDIA GTC
CoreWeave and Mistral AI are putting their infrastructure partnership on display at NVIDIA GTC, with a live conversation, a conference session and a hackathon run on CoreWeave Cloud. According to CoreWeave, the Paris-based startup trained its models on clusters of NVIDIA H100 Tensor Core GPUs hosted on its platform.
- CoreWeave says Mistral AI trained its models on NVIDIA H100 GPU clusters hosted on CoreWeave Cloud.
- Mistral Large is described as the world's second-ranking model available through an API, alongside GPT-4, based on common benchmarks.
- Arthur Mensch and CoreWeave's Max Hjelm will talk live at booth 1208 at 2:00 p.m. on Tuesday, March 19.
- Mensch leads a GTC session at 2:00 p.m. on Wednesday, March 20.
- Mistral AI will host its first Hackathon on CoreWeave Cloud in the Bay Area after the conference.
A fast-rising open-source lab
CoreWeave describes Mistral AI as a ten-month-old Paris-based startup that has reached the top tier of generative AI at a record-breaking pace. In recent months the company has released Mistral-7B and Mixtral-8x7B, La Plateforme, its API service, and Le Chat, a chatbot it calls "multilingual by design".
Its latest release is Mistral Large, a commercial model. According to the CoreWeave post, Mistral AI's results on commonly used benchmarks make it the world's second-ranking model available through an API, alongside GPT-4. CoreWeave says the model has unique reasoning capacities, is proficient in code and mathematics, and is fluent in French, English, German, Spanish and Italian. It can also process hundreds of pages of documents in a single call. These are claims drawn from CoreWeave's account and from benchmark figures it attributes to Mistral AI.
The company was founded by CEO Arthur Mensch, who left Google DeepMind, together with former classmates Timothée Lacroix and Guillaume Lample, who worked at Meta's Paris AI lab. Mistral AI offers both open models (Mistral-7B and Mixtral 8x7B) and commercial ones (Mistral Small, Mistral Medium and Mistral Large). Mensch is quoted: "Our ambition is to put frontier models in the hands of real-world application makers. Open source is a superb instrument to accelerate AI adoption."
What CoreWeave provides
CoreWeave says Mistral AI released Mistral-7B just a few months after the company was founded, thanks in part to quick access to a cluster of NVIDIA H100 Tensor Core GPUs on CoreWeave Cloud. Mensch describes the relationship in the post: "We worked hand-in-hand with members of CoreWeave's team to build the training clusters for our AI models". He adds that the performance of the clusters has been great, especially for reliability and stability.
CoreWeave's Max Hjelm, SVP of Revenue, is quoted as saying: "Everything in AI moves fast, but Mistral AI moves faster." CoreWeave also credits Mistral AI's cost-efficiency to its approach to model training and to the performance of its cluster on CoreWeave Cloud.
Events at NVIDIA GTC and the hackathon
At NVIDIA GTC, a global AI conference, Mensch will sit down with Hjelm for a live conversation at booth 1208 at 2:00 p.m. on Tuesday, March 19. Mensch will also lead a session at 2:00 p.m. on Wednesday, March 20. He plans to discuss Mistral AI's fast progress in open-source LLM training and the scientific lessons learned while training its first models, Mistral and Mixtral.
After the conference, Mistral AI will stay in the Bay Area over the weekend to host its first Hackathon, run on CoreWeave Cloud.
What it means
The post presents the collaboration as an example of how a young lab can scale quickly by relying on dedicated GPU cloud capacity rather than building its own infrastructure first. For developers, the hackathon offers a chance to work with Mistral AI's team on CoreWeave's platform. CoreWeave says the joint appearances highlight how the partnership supports cutting-edge model development for global enterprises.
