French AI lab Mistralhas released it’s newest model, Mistral Large 4, to the public—claiming it has chased its Chinese rivals to the frontier of open-weight AI.
Unlike most leading AI models, open-weight models are made freely available for anyone to download, run and modify. It is a class of AI models where the leading products have traditionally come from China-based companies.
The French startup says its new 1 trillion-parameter model, available on open-source platform Hugging Face from Thursday, is the best open model built in the West. Mistral says it narrowly tops all other open models on Artificial Analysis’ Coding Agent Index, an independent benchmark that measures how well AI agents complete full software engineering tasks.
According to benchmark figures provided by Mistral, the new model ties with China’s GLM-5.3, previously considered to be the best open model in the world, on Terminal-Bench 4.0, a test of coding agents. The company said that additional reinforcement learning training it has conducted since releasing a preview version of Large 4 on Oct. 6 has allowed the model’s performance on Terminal Bench 4.0 to jump from 28% accuracy to 41% accuracy. Proprietary models from companies like OpenAI and Anthropic, which can only be accessed through an API, remain ahead in overall capabilities. For instance, Anthropic’s Mythos 5.1 scored 60.9% on the same benchmark.
Mistral said, however, that Large 4 is also much more efficient than its Chinese open weight rivals, using 47% fewer tokens per task than GLM-5.3.
Large 4, which Mistral often refers to as “Le Chonk,” a humorous French wordplay based around the company’s cat mascot, was previously released as a limited preview with select partners. At that stage, Artificial Analysis rated it roughly on par with OpenAI’s GPT-6 Luna and DeepSeek’s V4.1 Flash, but found it cost more than four times as much per task as open models of similar intelligence. Pierre Stock, Mistral’s vice president of science, attributed the fresh gains the company is now reporting for the full release of Large 4 to three more weeks of reinforcement learning, a training method in which a model gets feedback on its attempts at tasks and learns to favor the approaches that work.
Stock called the release “a pivotal moment in the global AI race” and “a turning point for Europe,” arguing the continent can now build an ecosystem around AI foundation models, “not only the application layer.” He said the model was trained on Mistral’s own 10-megawatt computing cluster in France, “a fraction of what all the competitors use.”
Stock said the model was now “the best open-weight model on cyber in particular and coding.” He also suggested that Mistral’s model should be the defualt for companies and questioned whether Chinese open models can be trusted. Chinese models may be subject to future export controls, he said. He also said that a model’s behavior could be tweaked to exfiltrate secrets from a company’s systems.
Open-weight models are gaining ground with businesses who are attracted by the lower cost of these models and the greater ability for control over the data they process. Stock said “business continuity” is the top request from customers in finance, manufacturing and other regulated industries, such as healthcare, who worry that access to closed models “can be cut off.”
That risk was most keenly felt in June, when a U.S. export-control directive forced Anthropic to suspend access to its Fable 5 and Mythos 5 models for nearly three weeks. Stock argued that once a model’s weights are public “it’s impossible to cut off access” in this way. In addition, he said open models were essential for cyber defense, as closed models may rate-limit users or refuse legitimate requests to probe their own systems for vulnerabilities or engage in anti-hacking activities because of a similarity with some offensive cyber tasks.
In July, after OpenAI models breached Hugging Face’s servers during testing, Hugging Face’s team tried to respond using Anthropic’s Opus and Fable models. Those models refused much of the work because of their safety guardrails, and the team switched to an open model from China’s Z.ai.
Mistral was founded in Paris in April 2023 by Arthur Mensch, a former Google DeepMind researcher, and Guillaume Lample and Timothée Lacroix, who both came from Meta. It has released open models since its first one, Mistral 7B, came out later that year. The company has struggled to keep pace with U.S. labs, with its models generally trailing the top American systems.
It also faces an enormous funding gap. Mistral has raised roughly $6 billion in three years, most recently a €3 billion Series D in September led by Samsung, which valued the company at more than €21 billion (about $24 billion). Anthropic was valued at $965 billion as of August.
The post Mistral says its new model, Large 4, achieves ‘frontier’ performance among open models for agentic coding and cyber defense appeared first on Fortune.
