Mistral Large 4 is the new flagship artificial intelligence model from French firm Mistral AI, released as a public preview and nicknamed “le Chonk”. Built to process both text and images, the model closes performance gaps seen in earlier Mistral releases while placing itself among the leading open-weight systems available today.
The launch follows a busy period for the sector, arriving after the recent rollout of Claude Opus 5.5 and OpenAI’s GPT-6 Sol and GPT-6 Luna. Mistral AI, known for its open-source and open-weight large language models, reports that Large 4 ranks among the best open-weight models in the world on aggregated metrics. The company says the model outperforms European and US rivals while catching up to Chinese competitors, who have dominated the open-weight segment.
Key capabilities of Mistral Large 4
According to Mistral Chief Scientist Guillaume Lample and VP of Science Pierre Stock, Large 4 was engineered with advanced cybersecurity and agentic capabilities, alongside improved handling of manufacturing and finance-related tasks.
On cybersecurity, the model was designed to excel at defensive work, including log inspection, code auditing and checking for newly published exploits or vulnerabilities. Mistral says it also offers offensive cybersecurity measures to identify and address weaknesses.
Agentic performance sits at the heart of the release. “Really, the core competency we’re looking at for ML4 is agentic capabilities,” said Stock. The model was built to support planning, research and complex enterprise tasks that would otherwise take human experts hours or days. Mistral states it reaches state-of-the-art performance on critical benchmarks such as SWE-bench and Terminal Bench, evaluation frameworks that test how well agents handle coding, analysis and pulling requests.
Grounding and multilingual reach
The model is also said to outperform competitors in aerial and satellite imagery analysis, with the ability to identify visual patterns and objects on maps, such as forest fires. Mistral says that on grounding capabilities it outperforms all existing models, including closed ones.
Grounding is a technique used in AI development to reduce hallucinations, when a language model generates inaccurate information. It anchors outputs in external resources such as documents and databases rather than relying solely on training data, helping keep information current while providing context for responses.
Mistral adds that the new model was trained on more than 160 languages, including every official language of the European Union, as well as Latin and non-Latin languages, enabling it to answer prompts natively.
Mistral Large 4 will roll out on 2 October.
