Trending
BitFuFu considers abandoning 80MW Ethiopian Bitcoin mine acquisition Anthropic files S-1 for IPO, with $417bn in compute commitments ZutaCore details new two-phase boiling technology for liquid-cooled data centers Sponsored: The extinction of data breaches: Degauss, destroy, recycle Sponsored: The next evolution of data center backup power DCD Studio: Rethinking thermal management for AI scale with Oakley Roberts, Trane Cadence unveils Nvidia-powered AI data center in Munich, Germany The Next Competitive Advantage in Drug Development Won’t Be a Better Molecule. It’ll Be Better Patient Selection. An education on transfer switch neutral options Data Centers Bring Boom Times, But Will Gains Last? CTIA, Big 3 US carriers warn that support for D2D in 5G bands could hurt network performance Architecting AI infrastructure Verizon signs $2bn AWS AI infrastructure deal easyJet to offer Starlink in-flight WiFi from next year Ecoblox partners with Rimac on modular data center solution

Mistral launches Mistral Large 4 with 1 trillion parameters

Mistral AI released Mistral Large 4 on Tuesday, a multimodal artificial intelligence model whose weights it plans to make available in three weeks after safety testing.. The French AI lab’s model, known as ML4 and nicknamed “Le Chonk,” has 1 trillion parameters. It is not yet open-weight and is initially accessible only through a public guardrail endpoint..

“In the meantime, we’ll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks,” Mistral Vice President of Science Pierre Stock told TechCrunch.. Security concerns have increased in recent months, particularly among the enterprises and institutions that form Mistral’s core customer base.

Stock said open-weight models are easier to audit.. Mistral is positioning ML4 as an alternative to closed models that can be unplugged and open models that are often developed in China. The release aims to surpass American and Chinese rivals, following what French President Macron described as “a third way in AI.”.

Benchmark results are still pending. Stock said Mistral hopes ML4 will rank first among open-weight models, particularly, though not exclusively, outside China.. Focused training could also allow ML4 to outperform closed models in specialized areas relevant to customers where multimodal capabilities add value, Stock said.

That remains an ambition rather than a result established by the pending benchmarks.. ML4 was trained entirely on Mistral’s own computing infrastructure, using 4,000 Nvidia graphics processing units. Stock said that was “two to three times less than our Chinese competitors, and significantly less than the closed source competitors.”.

Mistral previously said its decision to host Chinese models did not represent a pivot toward becoming merely an inference provider. With ML4, the company believes it should continue to be considered a frontier AI lab.. According to Stock, ML4’s optimized use cases include cybersecurity, finance and chip design.

Chip design is central to the businesses of two of Mistral’s main backers, Dutch company ASML and Samsung.. ASML led Mistral’s Series C funding round. Samsung led its Series D last month at a valuation of 21 billion euros, approximately $24.39 billion..

Featured image credit. Tags: Mistral

 

Join the conversation

Your email address will not be published. Required fields are marked *