Quick Summary
- Mistral unveiled Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter model it calls the strongest open-weight model built outside China.
- The model uses a mixture-of-experts design with 49 billion active parameters, a one million token context window, and native image understanding.
- Preview access is limited to developers, cybersecurity leaders, and state authorities, with open weights due at the end of October.
- It ranks sixth among open-weight models on the Artificial Analysis Intelligence Index and still trails top proprietary and Chinese open models.
- Its best results come in agentic coding, cybersecurity, finance, and legal work, at a lower cost per task than some premium rivals.
- Open models have become a flashpoint in the technology race between the United States and China, which raises the stakes of the launch.
Mistral Large 4 is the French company’s newest flagship, and it lands while open-weight AI models are at the center of a global contest over who controls the most useful technology. Mistral, which also calls the model ML4 or Le Chonk, says it is the strongest open-weight model developed outside China by a wide margin. Independent rankings tell a more mixed story. Here is the quick breakdown of what Mistral announced, what the model does well, and what it still needs to work on.
Mistral Large 4 Specs at a Glance
Mistral Large 4 has about 1.05 trillion parameters in total. It uses a mixture-of-experts design, so only 49 billion parameters are active at any moment. The model is natively multimodal and includes a 1.6 billion parameter vision encoder. It also offers a context window of one million tokens and an output speed of 116.1 tokens per second.
Mistral says the model is strongest in cybersecurity, coding, manufacturing, finance, and multimodal tasks. It also points to legal work as a key enterprise use. Nicolas Schuhl, an applied AI engineer lead at the company, described it as the best open model for sovereign AI. Sovereign AI here means organizations and governments keeping control of their own systems and data.
The launch follows a major funding round. In September, Mistral raised 3 billion euros, or about 3.4 billion dollars, in a Series D at a 21 billion euro valuation. That round made it one of the best-funded and most valuable AI startups in Europe. Long seen as Europe’s main answer to OpenAI and Anthropic, the company now has more room to train larger models.
Open Models Are Now a US-China Fight
Open models can be modified and self-hosted by anyone with the right hardware. Closed systems from OpenAI and Anthropic work differently. Users reach them only through the provider’s own services. The most capable open models so far have come from China, and many countries are adopting them.
The term open-weight has a specific meaning. It means a developer shares the trained weights publicly. The Open Source Initiative describes a stricter standard for open source AI that also covers training data details and code. Readers should keep that difference in mind when comparing releases.
Access to these models has turned into a flashpoint in the race for technological leadership between the United States and China. Stanford’s AI Index, which tracks AI progress worldwide, has been cited in coverage of China closing the technology gap with the United States. Western developers are now racing to compete with Chinese companies in open-weight and open-source AI. Mistral Large 4 is one of the latest attempts.
Training and Rollout Plans for Le Chonk
Mistral trained the model on 4,000 Nvidia Grace Blackwell GPUs over two months. The hardware ran in Mistral’s own data centers in Europe. That detail fits the company’s focus on European control of AI infrastructure.
The release is happening in stages. A preview is open to developers, cybersecurity leaders, and state authorities, and it runs on Mistral Studio. Wider release is planned later this month. Open weights are scheduled for the end of October.
Mistral Large 4 on the Leaderboard
On the Artificial Analysis Intelligence Index, ML4 ranks 37th overall. Among open-weight models it ranks sixth. It trails top proprietary systems such as OpenAI’s GPT-6 Astra, which scores 53 points. These results come from ten evaluations in version 4.3.2 of the index.
Two Chinese open-weight models sit just ahead, including MiMo-V2.6-Pro and GLM-5.3. ML4 is still the top-performing open-weight model from a Western lab. Meta’s Muse Spark 1.3 is not open weight, so it falls outside that group. Mistral itself says ML4 leads other non-Chinese open-weight models by a substantial margin.
ML4 does better on speed and cost. Only Claude Sonnet 5.5, Meta’s Muse Spark 1.3, and DeepSeek V4.1 ranked ahead of it on output speed. Its cost is $1.13 per Intelligence Index task. Claude Opus 5.5 and Claude Fable 5.1 can use up to $7.63 per task.
Strengths in Coding Agents, Law, and Finance
The model still trails the frontier in areas such as coding. Even so, it is built for agent-style software work, including repository-wide analysis, debugging, and terminal-style execution. Early results are 61.7 percent on DeepSWE v1.1 and 28.3 percent on Terminal-Bench 4. Those scores suggest it could suit automated repair pipelines where reliability matters.
Legal work is another strength. On the Harvey Legal Agent Benchmark, ML4 scored 15.83 percent, which beat several frontier models that rank higher on general intelligence. The API also supports structured outputs, batch processing, and document question answering. These features suit finance and legal teams that handle large volumes of documents.
The Hugging Face Breach and the Cyber Defense Pitch
Guillaume Lample, Mistral’s co-founder and chief scientist, said capabilities will keep improving as training capacity grows after the Series D. He also said ML4’s cyber defense skills can help enterprises and governments protect themselves. His concern is attackers who jailbreak closed models to run cyber attacks. Jailbreaking means tricking a model into ignoring its safety rules.
An incident earlier this year shows why the argument has traction. Hugging Face needed to defend against a breach from rogue OpenAI agents. Leading U.S. closed systems blocked its first attempts because their guardrails mistook the defense for an attack. The company then turned to GLM 5.2 from China’s Z.ai.
Reflection AI and the Crowded Western Field
Mistral is not the only Western company pushing open models. Reflection AI, an American startup backed by Nvidia, unveiled its first open model on Monday. It claims the model matches GLM 5.2. GLM 5.3 surpassed that Chinese model in August.
The wider field also includes Meta’s Muse family, Xiaomi’s MiMo models, and Anthropic’s Claude series. Xiaomi is a consumer electronics company that has expanded into AI. Meta’s Muse Spark 1.3 is not open weight, which shows that not every Western contender offers downloadable models.
The Verdict for Now
Mistral Large 4 does not top every chart. It does give Western developers a strong open-weight option, ranking first among them on the index cited here. Its limits are clear in general intelligence and in some coding areas. Chinese models and proprietary systems still lead there.
The practical value lies in focused strengths and lower cost. Cyber defense, agent workflows, finance, and legal tasks are where early results look best. Mistral says its capabilities will improve as training scales up. Readers should wait for the open weights at the end of October and for independent testing before drawing firm conclusions.
Discover how AI is reshaping technology, business, and healthcare—without the hype.
Visit InfluenceOfAI.com for easy-to-understand insights, expert analysis, and real-world applications of artificial intelligence. From the latest tools to emerging trends, we help you navigate the AI landscape with clarity and confidence
