Mistral AI has launched Mistral Large 4, nicknamed Le Chonk, a 1 trillion parameter model with 49 billion active parameters, trained natively with multimodal capabilities . The Paris-based company is positioning the release as the strongest open-weight foundation model to come out of the US or Europe on aggregated benchmarks, with specific emphasis on state-of-the-art performance in critical workloads such as cybersecurity, manufacturing, and finance, and surpassing closed-source frontier models in visual grounding . For founders and operators building on sovereign infrastructure, this gives European teams a frontier-grade alternative that does not require sending data to US closed-source APIs.
The launch matters right now because the model is available via API, with the open-weight version planned for release at the end of October , letting teams prototype against the hosted endpoint and plan self-hosted deployments on the same model family. The combination of a mixture-of-experts architecture (1T total, 49B active) and native multimodality means buyers evaluating vision plus text workloads, document understanding, or agentic pipelines can benchmark it directly against GPT-class and Claude-class systems without waiting on a separate vision model.
Mistral AI was founded on 28 April 2023 by Arthur Mensch, Guillaume Lample, and Timothée Lacroix , who continue to serve as CEO, Chief Scientist, and CTO respectively. The company has become the anchor of the "sovereign AI" conversation in Europe, and Le Chonk is the clearest signal yet of how it plans to compete with US labs on raw capability while keeping weights open for enterprise and government customers that require on-premise control.
Comments (12)
Sign in to join the discussion.
P
Priya Varadarajan7d ago
Naming a 1T parameter model Le Chonk is the kind of unserious branding that only works if the benchmarks actually hold up. Luckily for you, baguette boy appears to deliver.
R
Rohit Sankaran7d ago
Aggregated benchmarks is doing an enormous amount of work in that sentence. Which eval suite, which temperature, and who picked the subset?
K
Kjeld B.7d ago
The sovereignty angle hits different when half of Europe is still arguing about whether to allow a chatbot in the ministry. Good timing on the drop.
N
Nnamdi Okafor7d ago
Cool, but does Le Chonk come with SOC2 Type II, a signed DPA, and a sales rep who returns emails before Q3? Asking for my entire security review board.
M
mila edits7d ago
The launch tweet has no video and no chart. One billion parameters per word of marketing copy. Respect the restraint or fire the creative director, I cannot tell which.
K
Katarina Vuković7d ago
Been quietly bullish on the Paris stack for a while. The cyber defense wedge is the smart one, nobody wants to send their SOC logs to a Bay Area startup.
H
Haruki Oshima7d ago
A model you can host is worth two you cannot audit.
Y
Yewande A.7d ago
Dumb question but if it is natively multimodal, why does the landing page only show text examples? First impression matters and I was looking for the image demo.
O
Oluwaseun T.7d ago
Tried Le Chat right after reading this and the model picker still buries the new one three clicks deep. Ship the launch into the product, not just the tweet.
D
Dmytro Hrytsenko7d ago
49B active on 1T total means you still need serious hardware to serve this at latency. Curious what the recommended deployment footprint looks like for actual customers.
B
Beatriz Camargo7d ago
Fourth frontier model launch I've seen this month and the only one that didn't open with a cinematic drone shot of a GPU rack. Points for that alone.
F
Freya Lindqvist7d ago
If anyone on the Mistral post-training team is open to chatting, I have a former DeepMind RLHF lead in Zurich who would love a coffee. Le Chonk needs friends.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Mistral AI
#1 in Sovereign AI
#1 in Sovereign AI
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Mistral AI has launched Mistral Large 4, nicknamed Le Chonk, a 1 trillion parameter model with 49 billion active parameters, trained natively with multimodal capabilities . The Paris-based company is positioning the release as the strongest open-weight foundation model to come out of the US or Europe on aggregated benchmarks, with specific emphasis on state-of-the-art performance in critical workloads such as cybersecurity, manufacturing, and finance, and surpassing closed-source frontier models in visual grounding . For founders and operators building on sovereign infrastructure, this gives European teams a frontier-grade alternative that does not require sending data to US closed-source APIs. The launch matters right now because the model is available via API, with the open-weight version planned for release at the end of October , letting teams prototype against the hosted endpoint and plan self-hosted deployments on the same model family. The combination of a mixture-of-experts architecture (1T total, 49B active) and native multimodality means buyers evaluating vision plus text workloads, document understanding, or agentic pipelines can benchmark it directly against GPT-class and Claude-class systems without waiting on a separate vision model. Mistral AI was founded on 28 April 2023 by Arthur Mensch, Guillaume Lample, and Timothée Lacroix , who continue to serve as CEO, Chief Scientist, and CTO respectively. The company has become the anchor of the "sovereign AI" conversation in Europe, and Le Chonk is the clearest signal yet of how it plans to compete with US labs on raw capability while keeping weights open for enterprise and government customers that require on-premise control.
Comments (12)
Naming a 1T parameter model Le Chonk is the kind of unserious branding that only works if the benchmarks actually hold up. Luckily for you, baguette boy appears to deliver.
Aggregated benchmarks is doing an enormous amount of work in that sentence. Which eval suite, which temperature, and who picked the subset?
The sovereignty angle hits different when half of Europe is still arguing about whether to allow a chatbot in the ministry. Good timing on the drop.
Cool, but does Le Chonk come with SOC2 Type II, a signed DPA, and a sales rep who returns emails before Q3? Asking for my entire security review board.
The launch tweet has no video and no chart. One billion parameters per word of marketing copy. Respect the restraint or fire the creative director, I cannot tell which.
Been quietly bullish on the Paris stack for a while. The cyber defense wedge is the smart one, nobody wants to send their SOC logs to a Bay Area startup.
A model you can host is worth two you cannot audit.
Dumb question but if it is natively multimodal, why does the landing page only show text examples? First impression matters and I was looking for the image demo.
Tried Le Chat right after reading this and the model picker still buries the new one three clicks deep. Ship the launch into the product, not just the tweet.
49B active on 1T total means you still need serious hardware to serve this at latency. Curious what the recommended deployment footprint looks like for actual customers.
Fourth frontier model launch I've seen this month and the only one that didn't open with a cinematic drone shot of a GPU rack. Points for that alone.
If anyone on the Mistral post-training team is open to chatting, I have a former DeepMind RLHF lead in Zurich who would love a coffee. Le Chonk needs friends.