Mistral Large 4 Preview: European Open Weights, Cyber Benchmarks and Setup Friction

Mistral has introduced a public preview of Mistral Large 4, nicknamed Le Chonk. It is a European model trained from scratch in the trillion-parameter class. I am excited to see another region competing in open-weight AI, although the weights were not yet released when I recorded this.

The model combines instruction following, reasoning and agentic capabilities with multimodal input. It uses a mixture-of-experts architecture: the total model is large, but only a subset of parameters is active for a given task. Mistral states that it has about one trillion total parameters and 49 billion active parameters.

The preview’s quoted API prices are $1.36 per million input tokens and $4.18 per million output tokens. I consider those inexpensive, though price and capability must be compared with the alternatives. Proprietary models have also improved quickly and become cheaper and faster.

Open weights offer a different trade-off: organizations can potentially host and control a model themselves, including how data is handled. Mistral said weights were planned for release by the end of the month. During the preview it was testing cyber capabilities with vetted partners and authorities under different access conditions.

On Mistral’s displayed DeepSWE 1.1 comparison, Large 4 Preview scores 62, behind Kimi K3 at 68 and ahead of several other listed open models. The company’s charts also show competitive Terminal-Bench, cyber and agent results. Rankings depend on the models and configurations included in each chart.

The displayed Artificial Analysis cyber index puts Mistral Large 4 at 50, tied with GLM 5.3 Flash among the shown models. On AutomationBench it scores 59.9 versus 62.2 for GLM 5.3. I see a major improvement from the earlier Mistral model, although it does not lead every test.

Mistral also highlights finance and legal-agent performance. I note that a legal comparison omits Grok 4.7, which I recall reporting 19.6 on that benchmark. When comparing company charts, it matters which competitors and versions are included.

Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data center, and serves the preview on the same infrastructure. I value this as a question of AI sovereignty: competition between regions and the ability to run models independently matter alongside competition between companies.

At the time, the preview was available through an API rather than the Le Chat application. I tried connecting it to OpenCode and found the setup difficult. It produced so many reasoning tokens that it hit the configured context limit and stopped without useful feedback until I debugged it.

This is a broader adoption problem. Most users do not want to manage context limits, API keys and agent harness settings. For open models to reach a general audience, I think they need a plug-and-play experience as simple as leading proprietary coding tools.

Open models can already work well for an organization with the expertise and scale to route suitable tasks to a model it controls. I distinguish token volume from revenue: in my account, open models can represent a large share of usage while proprietary providers capture substantial value. My own experience shows that making this work still takes effort.

Mistral emphasizes cybersecurity, including a strong CyberGym result. Its description places the model among the strongest on the Artificial Analysis cyber index. Those are benchmark results for particular evaluations, not a guarantee that the model will find or fix every real security flaw.

On the displayed Artificial Analysis intelligence index, Large 4 Preview scores 38 and appears below several proprietary and open models. I admire the European effort but do not take that chart to mean it is the strongest general-purpose model. Different evaluations and model configurations tell different parts of the story.

The price-versus-intelligence chart makes the trade-off clear in my view: some proprietary alternatives offer stronger general performance at competitive prices. Open weights can still provide deployment control and give a community room to optimize or adapt the model for narrower tasks.

The displayed context window is 500,000 tokens, below the one million shown for several competitors. The speed chart puts the preview around 116 output tokens per second in that measurement. Actual speed depends on how and where a model is served.

Mistral’s CyberGym chart shows a score of 82, leading the open models included there. Berman notes that major proprietary models were absent from that particular chart. The result is encouraging, but the selection matters when interpreting a ranking.

I tested the model by asking for a Rubik’s Cube simulator. At high reasoning effort it used roughly 32,000 tokens and hit a configured limit. Increasing the limit and turning off reasoning improved progress, but the debugging reinforced my concern that setup is too hard for most users.

The first cube looked good, but the scramble animation did not move. After another iteration, the sides turned while colors disappeared, and the final solve appeared to reset colors. I count the test as a failure.

I attribute some of the difficulty to the interaction between model and agent harness rather than assuming the model alone caused it. A well-optimized harness can make a large difference; Cursor is an example I have found easy to use with other companies’ models. I had less success with OpenCode in this test.

Despite the rough edges, I welcome a competitive model trained in Europe and the prospect of open weights. The model needs a more polished companion workflow, and I still need to test it with a better-suited harness. More regional competition and more ways to deploy AI are good developments in my view.