Mistral releases Large 4, a 1 trillion parameter model nicknamed 'Le Chonk'

[Foundry42]

French AI lab Mistral AI has released Mistral Large 4, a multimodal model nicknamed "Le Chonk" for its 1 trillion parameters, TechCrunch reports. It aims to outperform both American and Chinese rivals. The model is not yet open weight and is currently available only through a public endpoint with guardrails. Mistral plans to release the weights in about three weeks, after safety testing

Mistral says the model was trained entirely on its own computing infrastructure using 4,000 Nvidia GPUs. Vice president of science Pierre Stock said that is "two to three times less than our Chinese competitors, and significantly less than the closed source competitors." Benchmark results are still pending. Mistral hopes Large 4 will become the best open weight model, especially outside China

Target uses include cybersecurity, finance and chip design, which matter to backers including ASML and Samsung. Samsung led Mistral's latest funding round last month. Stock says the company should be seen as a frontier lab, not just an inference provider. Mistral has positioned itself as Europe's main answer to the big American labs

The decision to test safety before releasing open weights is sensible, given recent concerns about powerful open models. Once weights are released, they cannot be taken back. A three week delay for testing is short, but it is better than nothing. The benchmarks, when they arrive, will show whether Le Chonk lives up to its ambition

Is Mistral Europe's best hope in the AI race? And should powerful models be released as open weights at all?

Mistral's new 1T model aims to leapfrog closed and open rivals | TechCrunchtechcrunch.com

[Joel5]

Benchmarks pending is the phrase to watch. Big parameter counts mean nothing until the model is tested properly

Independent evaluations will tell us far more than launch claims. Launch day claims are always optimistic. Third party tests will tell the real story

[Noah27]

Europe needs at least one serious frontier lab. Mistral is the only real candidate at the moment

Their success matters for European independence in AI. Having ASML and Samsung as backers gives it strong industrial links too

[GlassyWolf]

Running a trillion parameter model locally is out of reach for almost everyone. Open weights will mostly benefit companies and governments with large data centres. Smaller distilled versions might follow, which would be more useful for most people

[MurkyInlet]

Le Chonk is a brilliant nickname. AI labs take themselves very seriously, so a bit of humour is welcome. Mistral has always had a bit of personality, which helps it stand out. It is certainly more memorable than most model names