French company Mistral AI unveiled on October 6, 2026 a preview of Mistral Large 4, a one-trillion-parameter model that the Paris-based company says is the most powerful open-weight artificial intelligence system outside China. From that day, developers can use it through Mistral's API, and the weights, which let anyone download and run it on their own servers, will be available on October 27.
According to The Next Web, Mistral will spend the next three weeks testing the model under real-world conditions. During that period, developers, cybersecurity firms and government agencies will receive a version with fewer restrictions and more cyber features than the one on the public API. Internally, the model is nicknamed “Le Chonk,” and it processes both text and images.
In terms of architecture, Large 4 uses only 49 billion of its parameters for each task. It was trained from scratch over two months on about 4,000 NVIDIA Grace Blackwell GPUs in Mistral's European data centers, which now also serve the preview. The company estimates that training drew around 10 megawatts of power. The reinforcement learning phase is still ongoing, and Mistral says the gains have not yet begun to slow.
The published results are preliminary, and the company warns they may change before the weights are released. On DeepSWE, an agentic coding test built on real software engineering tasks, Large 4 scored 62%, narrowly ahead of Zhipu's GLM-5.3 (61%) and DeepSeek-V4-Pro (57%), and well clear of Qwen 3.8 Max (51%) and Reflection AI's Beam (44%).
In finance, the model reached 67% on FinWorkBench, level with DeepSeek-V4-Pro and ahead of GLM-5.3 (65%), and almost double the 37% of Mistral Medium 3.5. On Harvey's Legal Agent benchmark it scored 15%, compared with 13% for Kimi K3 and 5% for OpenAI's GPT-6 Astra, although the low scores across the board show how far AI still is from handling legal work on its own.
Vision is where Mistral stands out most. On locating objects in remote-sensing images, Large 4 scored 73% on DIOR-RSVG, compared with 68% for GPT-6 Astra, and both models tied at 42% on Dense200. Mistral also claims it beats Anthropic's Fable 5.1 on these tasks but has not published the numbers for that comparison. The company envisions uses such as assessing storm damage in aerial photos, inspecting power lines and monitoring crops.
The model can also convert technical drawings into CAD models and, according to Mistral, performs very well on semiconductor-related tests, which points to industrial customers. Cybersecurity, however, sits at the center of the launch. The most powerful models from American labs are closed: they can only be rented through the provider's servers, and the provider decides what the system will and will not do.
Once Large 4's weights are published, a bank or a government will be able to download it, run it on its own machines and use it to protect its networks without asking anyone's permission. For security teams, which need to scan their systems daily for new weaknesses, closed models are often too expensive, too restricted or simply refuse the request. Mistral says Large 4 beats the best open models from Kimi, DeepSeek and Meta on cyber tasks, though it has not shared figures to back that up.
Guillaume Lample, Mistral's co-founder and chief scientist, summed it up: “The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks.” The company adds a political argument: companies and governments should not depend on a provider that could switch off their tools at any moment.
That sovereignty idea shapes how the model will be sold. Customers will be able to run it on their own servers or use it through Mistral's API in the region of their choice, including a European region where data is protected by EU law. Large 4 was trained on more than 160 languages, including every official language of the European Union and others such as Korean.
Funding comes from the €3 billion round Mistral closed in September; Large 4 is the first model to come out of that investment, which will also pay for more computing capacity due online in the first half of 2027. Mistral works with more than 125 large businesses, including Airbus, ASML and HSBC, and some of them helped train Large 4 using the same tools the company sells through its Forge platform.
Before October 27, Mistral promises to publish more details on how the model was built, more benchmark results, how it was refined after its main training and how it was safety-tested. Afterwards, the company will use Large 4 as the base for a new family of specialized models. Its co-founder and chief executive, Arthur Mensch, argued at a conference in Abu Dhabi that “the narrative that Europe cannot compete is something that is not true.”
Ultimately, Large 4 is Mistral's most ambitious bet to date, and it arrives amid competition with Reflection, which unveiled Beam a day earlier, and with Chinese labs. If the weights ship on time and the numbers hold up, Europe will have a top-tier open model geared toward cyber defense, industry and technological sovereignty.
