started · updated
Mistral AI unveils Mistral Large 4 trillion-parameter AI model
Mistral AI has unveiled Mistral Large 4, a massive 1-trillion-parameter multimodal model nicknamed “Le Chonk.” Utilizing a mixture-of-experts (MoE) architecture, the system activates 49 billion parameters per query to balance high performance with manageable computational costs. The model is designed to compete with leading proprietary systems from the United States and open-weight models from China, specifically targeting sectors such as cybersecurity, coding, finance, and manufacturing.
In cybersecurity benchmarks, Mistral claims the model is among the top five globally, scoring 82% in tests for reproducing and patching open-source software vulnerabilities. This capability is a central pillar of Mistral’s strategy to provide “sovereign AI,” allowing governments and enterprises to maintain control over their data and defense tools without relying on closed-source vendors.
The model was trained from scratch over two months using approximately 3,800 to 4,000 Nvidia Grace Blackwell GPUs in Mistral’s European data centers. It supports over 160 languages, including all official languages of the European Union. While a public preview is currently available via API, the full model weights are scheduled for release on October 27, 2026, following security testing with trusted partners and government authorities.
Entities
ASML · Arthur Mensch · Artificial Analysis · Artur Mensch · Guillaume Lample · Mistral · Mistral AI · Mistral Large 4 · Nvidia · Samsung
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 19 SOURCES] The model uses a mixture-of-experts architecture that activates 49 billion parameters per query. unwire.pro · bitnewsbot.com · the-decoder.de · borncity.com · www.generation-nt.com · +14 more
- [● 24 SOURCES] The model weights are scheduled for release on October 27, 2026. unwire.pro · bitnewsbot.com · imprensapublica.com.br · valor.globo.com · www.developpez.net · +19 more
- [● 11 SOURCES] Mistral raised €3 billion in a Series D funding round led by Samsung. unwire.pro · bitnewsbot.com · imprensapublica.com.br · srnnews.com · www.constellationr.com · +6 more
- [● 6 SOURCES] The model's input cost is $1.36 per million tokens and output cost is $4.18 per million tokens. bitnewsbot.com · borncity.com · www.constellationr.com · www.itmedia.co.jp · tech-noisy.com · +1 more
- [● 4 SOURCES] Mistral Large 4 scored 82% in a test for reproducing and patching open-source software vulnerabilities. unwire.pro · www.diarioestrategia.cl · siliconangle.com · brusselssignal.eu
- [● 9 SOURCES] The model supports more than 160 languages. unwire.pro · imprensapublica.com.br · www.generation-nt.com · borncity.com · ihal.it · +4 more
- [● 29 SOURCES] Mistral Large 4 is a 1-trillion-parameter multimodal model. espiganoticias.net · unwire.pro · bitnewsbot.com · imprensapublica.com.br · valor.globo.com · +24 more
- [● 14 SOURCES] The model was trained using approximately 4,000 Nvidia Grace Blackwell GPUs in European data centers. espiganoticias.net · unwire.pro · www.developpez.net · www.diarioestrategia.cl · borncity.com · +9 more