Mistral’s Large 4 Makes Open Weights Practical for Enterprise AI

Mistral has released Large 4 to everyone. The trillion-parameter model, also known as “Le Chonk,” is now available as downloadable open weights under the Mistral Ecosystem Open-Weight License, following a limited public preview introduced on October 6.

The final release adds three weeks of reinforcement learning and feedback from customers and security partners. Mistral says that work pushed Large 4’s Terminal-Bench 4.0 accuracy from 28% to 41%, while independent testing shows the larger gains came from lower costs, fewer fabricated answers and better efficiency rather than a dramatic leap in raw capability.

Artificial Analysis gives the released model 26 points on Intelligence Index v5 at its high reasoning setting, up from 25 for the preview. That puts Large 4 one point above Muse Spark Open at xhigh and makes it the highest-scoring open-weight model developed outside China in the evaluation, although it remains six points behind DeepSeek V4.1 Flash 1017 and GLM-5.3 at their maximum reasoning settings.

Better economics, not just bigger numbers

The cost improvement is easier to appreciate than another benchmark trophy. Artificial Analysis calculates that Large 4 costs $2.06 per Intelligence Index task, down from $5.54 for the preview—a 63% reduction that measures the released checkpoint rather than projecting savings from token prices.

Mistral’s final API rates are $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. Those figures are roughly half the earlier $1.36, $0.14 and $4.18 rates, while the model’s 1 million-token context capacity gives it room for large codebases and extended work sessions.

Large 4 also uses 47% fewer tokens per task than GLM-5.3, according to Mistral. On Artificial Analysis’s Coding Agent Index, it ranks second among current open-weight entries, one point behind GLM-5.3 and two points above DeepSeek; the comparison includes the model and the agent software directing its tool calls, so the leaderboard is not measuring model weights alone.

Mistral says Large 4 is the best open model for agentic coding and cyber defense. The company points to tasks such as creating patches and reproducing vulnerabilities, while saying the model refuses malicious cyber requests more often than other open models.

Open weights with a business boundary

The release offers 1.05 trillion total parameters and 52 billion active parameters, compared with the 49 billion active parameters cited for the preview. Mistral’s Hugging Face listing explains that active parameters measure activity per token, while the total also includes embeddings and output layers; the final model used 3,800 Nvidia Grace Blackwell GPUs for training instead of the approximately 4,000 reported initially.

Organizations can download and serve the weights with vLLM or SGLang. Deployment options include Mistral Studio and its API, OpenRouter, Vibe Code, Microsoft Foundry with private Azure deployment options, plus Nvidia-hosted endpoints and an optimized NIM container.

The license grants broad rights to run, copy, modify, further train, distribute, sublicense and sell the model and its derivatives, provided copyright and permission notices accompany copies or substantial portions. It contains no revenue threshold—a welcome break from licensing documents that treat company size as a moral category.

Section 2 requires a separate commercial license before commercial use if the licensee or an affiliate operates a model-as-a-service business. That covers hosted inference or post-training through an API, hosted endpoint or another method when third parties retain meaningful control over inputs, parameters or training data.

The license excludes end-user products where model capabilities sit solely inside specific features or harnesses, along with services that relay requests to models hosted by others. Internal use that does not expose the model or its derivatives to third parties is also exempt, as is use through Mistral’s official products—so an enterprise can run Large 4 internally without triggering Section 2’s requirement.

That distinction concerns business activity and access, not revenue. Downloading the weights does not make hosted services or the infrastructure needed for self-hosting free, and operators offering customers the specified control still face the separate-license requirement.

Control is the actual pitch

Marjorie Janiewicz, Mistral’s chief revenue officer, described the enterprise case as a question of control: “I don’t want to get disconnected. I want to own my model.” She added, “Second, I want to customize it to ensure that I derive value. And then third, it’s business continuity.”

Pierre Stock, Mistral’s vice president of science, called the release “a pivotal moment in the global AI race” and “a turning point for Europe.” He said Europe can build an ecosystem around foundation models, “not only the application layer,” using Mistral’s own 10-megawatt computing cluster in France—“a fraction of what all the competitors use.”

Stock called Large 4 “the best open-weight model on cyber in particular and coding,” while arguing that access cannot be cut off in the way closed models can. Mistral’s Arthur Mensch framed the broader goal as making AI “a tool for empowerment, not dominance,” with “state-of-the-art performance” at “a fraction of the compute and capital committed by closed labs.”

That argument matters most to finance, manufacturing, healthcare and other regulated industries worried about access being cut off. Open weights do not erase the cost or difficulty of deployment, but they give organizations a model they can retain, customize and run inside their own systems—three things a permanent API dependency does not provide.

Stay connected via Google News

Follow us for the latest AI News