With Large 4, Mistral Tries to Put Europe Back in the A.I. Race

Set Trending Topics as a preferred source on Google.

After a long silence, a comeback has to be loud: Mistral has just unveiled Mistral Large 4, and with it the French company wants to prove that Europe is still in the A.I. race. The model is meant to close the gap with the best open-source models, including the Chinese ones. For now, though, there are no independent benchmarks based on the open weights that could confirm it.

The new model, called ML4 internally and nicknamed “le Chonk” with a wink, beats Z.ai’s GLM 5.2 in Mistral’s in-house benchmarks and is said to be “by far” the most powerful open model from Europe and the United States. Large 4 is a mixture-of-experts model with one trillion parameters, 49 billion of which are active at a time. Mistral says it can keep up with the best closed models while competing with open models three times its size. Mistral is starting with a preview through its API, priced at $1.36 (about €1.21) per million input tokens and $4.18 (about €3.71) per million output tokens, and the weights for download and self-hosting are set to follow at the end of October. The main target is companies that want to run the model in their own data centers or in a private cloud. It is too large for laptops and desktop computers. Large 4 was originally supposed to arrive in the summer. The launch slipped, its predecessor is already almost a year old, and in the meantime the French A.I. company had gone strikingly quiet.

What Large 4 Is Supposed to Do

Mistral has tailored the model to the needs of its enterprise customers in banking, logistics, defense and the public sector. The main new features, according to Mistral:

• Coding: Large 4 is designed to analyze existing software projects and add new features on its own. Mistral says the model closes the gap with the leading models in coding and is especially good at designing and testing semiconductors.
• Agents: The model is meant to carry out complex tasks over several hours and work with a range of tools without losing context. According to Guillaume Lample, Mistral’s chief scientist and a co-founder, it can take on work that has so far kept specialists busy for hours or days.
• Cybersecurity: Mistral says the model is equally strong on offense and defense. Customers wanted a European model to find security flaws and fend off attacks, said Pierre Stock, Mistral’s vice president of science. Mr. Lample told The Deep View that the open weights are the key argument: New security workflows require a lot of computing power, “so you cannot afford to be vulnerable to the fact that the model you are using to protect yourself might disappear one day, or might be too limited.” Because these capabilities can also be abused for attacks, Mistral says it monitors traffic through its API.
• Domain knowledge: Beyond coding and security, the model is supposed to perform particularly well on tasks in manufacturing, finance and geospatial analysis.
• Grounding: In visual grounding, meaning locating objects in images such as satellite and aerial photos, Mistral claims Large 4 “outperforms all existing models, including the closed ones.”
• Images and languages: Large 4 is multimodal and processes images as well as text. It was trained on more than 160 languages.

How Mistral Ranks Itself

In its launch blog post, Mistral chooses its words carefully: Large 4 is competitive with the strongest open models worldwide and significantly outperforms any open-weight model developed in the United States or Europe. In other words, Mistral does not claim a clear win over China’s best. On enterprise workloads such as cybersecurity, finance and law, however, it sees itself at the top of the open models, and in visual grounding even ahead of closed frontier models. Parts of the model are “basically stronger than China’s models from this summer,” Mr. Stock told Euronews. The key comparisons:

Against the Chinese competition, the margins are thin: On DeepSWE, Large 4 edges out GLM-5.3, and on FinWorkBench it ties with DeepSeek V4 Pro. In an internal comparison with GLM-5.3, expert annotators preferred Large 4 for CAD as well as math and physics, while the two were roughly even in finance and coding. In the blind coding evaluation, Large 4 came in second of five, behind Anthropic’s Claude Opus 5. The only other American models in the comparisons are Reflection AI’s recently unveiled Beam, which trails well behind in coding, and OpenAI’s closed GPT-6 Astra.

Mistral is at its most confident on cybersecurity. On the Artificial Analysis Cyber Index, an independent evaluation, Mistral says Large 4 ranks among the top five models globally and leads open-weight models developed outside China by a wide margin. On a test that asks a model to reproduce a real vulnerability and then patch it, it scores 82 percent, the highest of any model. Closed models such as Claude Opus 5.5 and GPT-6 Astra score near zero there because they refuse the task. Until the weights are released, security companies, vetted partners and state authorities are testing the model with reduced moderation.

A Model With Symbolic Weight

For Europe’s tech scene, the launch is more than a version bump. Mistral is seen as Europe’s most important counterweight to the American heavyweights, and every new flagship model is read as a gauge of whether European companies can remain even remotely competitive. Mistral has recently won major industrial customers such as ASML, Samsung and Airbus, along with partnerships with countries including France and Luxembourg.

The pressure had been mounting. Competitors from China and the United States had left Mistral’s earlier models behind. The A.I. researcher Niels Rogge criticized on LinkedIn that Large 3.5 was built on an “outdated architecture” from Meta and cost more than stronger models from DeepSeek. Some observers already saw Mistral drifting away from cutting-edge research toward becoming a pure service provider for businesses. Data obtained by the industry publication Sifted backs that up: In Mistral’s early days, more than 80 percent of its technical hires were A.I. researchers and research engineers. In the first half of 2026, the research team made up only 29 percent of the technical staff.

Mr. Lample pushes back: “Our closed-source competitors are at least four years older than us, but we now have models that are really strong. We have essentially closed that gap and we are accelerating.” As evidence of its own research strength, Mistral points out that Large 4 was trained with reinforcement learning. Many open models, by contrast, are largely distilled from other models.

The Strategy Stays the Same: A Neocloud, Not Just a Model Maker

The new model does not change Mistral’s direction. The company keeps building out its neocloud, meaning its own computing infrastructure and an offering that goes beyond developing models. Large 4 was already trained in Mistral’s own data centers in Europe, on 3,800 Nvidia Grace Blackwell chips and, according to Mistral, at a fraction of what competitors spend. The preview runs on the same infrastructure, so the company supplies the model, the infrastructure and the computing power from a single source. To fund this shift, Mistral recently raised €3 billion at a valuation of €21 billion. According to Sifted, its total equity and debt financing now exceeds €6 billion. The money is mainly going into model development and more computing capacity, according to Johan Bergqvist, Mistral’s chief financial officer, to give companies that want to run their own models an alternative.

That is Mistral’s strongest selling point in Europe: independence from foreign providers. How delicate that dependence can be became clear in June, when Anthropic said the U.S. government had instructed it to block foreign nationals from accessing some of its most powerful models. “Customers want open models because they want to be sure that access won’t soon be cut off by an external power,” Mr. Stock said.

Consistent with that, Z.ai’s GLM 5.2 will remain available on Mistral’s platform, even though Mistral’s internal tests put its own model ahead. Mistral increasingly presents itself as a platform offering a range of models rather than a provider that relies solely on its own weights.

A Look at the Competition

The field is crowded. OpenAI, Anthropic and Google are locked in a tight race at the top, while Chinese providers such as Z.ai, Alibaba with Qwen, Moonshot AI with Kimi and DeepSeek are putting heavy pressure on the open-model market. That is exactly where Mistral wants to establish itself as the European alternative, among powerful models that companies can deploy flexibly.

The Caveat: In-House Tests

The comparisons with GLM 5.2 and GLM-5.3 come from Mistral’s own preliminary benchmarks. Claims like these can only be put into context once independent tests are available, and that will be possible at the earliest when the weights are released at the end of October. Large 4 is not yet listed in the Artificial Analysis Intelligence Index, which combines results from numerous benchmarks. The Mistral models already listed there trail far behind the open-weight leaders: Mistral Medium 3.5 scores 14 points and its predecessor Large 3 scores 9. GLM-5.3, the successor to GLM 5.2, reaches 45 points, and Moonshot AI’s Kimi K3 reaches 44. Large 4 would have to make a big leap to join that league. On top of that, Mistral says the reinforcement learning run behind the preview is still underway, so the numbers are likely to keep changing.

Large 5 Is Already in Pretraining

As Large 4 launches, Mistral is already working on its successor. Mistral Large 5 is reportedly already in pretraining. Whether Mistral can keep pace with its rivals will depend on how quickly Large 5 follows and how well it holds up in a direct comparison.