Mistral releases open weights for Large 4, its 1-trillion-parameter AI

Mistral released the open weights of Mistral Large 4 (ML4), its 1-trillion-parameter flagship nicknamed “Le Chonk”, on October 29, and organisations can now download it from Hugging Face. The most important change since the October 6 preview is that Mistral kept training the model for three more weeks: its score on the Terminal-Bench 4.0 coding test rose from 28% to 41%, on the company’s own figures.

At preview, Mistral promised open weights by the end of October, and it has delivered on that date. The French company now calls Large 4 “the best open agentic coding model in the world” and says it gives enterprises and governments “frontier AI cyber defence capabilities they can fully own and control”. The model remains available through Mistral’s API alongside the download.

What changed between the Large 4 preview and the open release

Mistral says it spent the three weeks since the preview continuing reinforcement learning and testing the model with customers and partners under real-world conditions. Reinforcement learning is a training method in which a model gets feedback on its attempts at tasks and learns to favour the approaches that work.

“ML4 OSS showcases our science team’s ability to innovate at speed. Building on the ML4 Preview, the open weight version continues reinforcement learning training and delivers substantial gains with no signs of saturation.” — Pierre Stock, Vice President of Science at Mistral

Stock added that within the same training run, the team “integrated customer feedback by introducing two additional reasoning modes, while making ML4 more efficient and customisable”. In practice, that means:

• Three reasoning modes. Users can now pick low, medium or high reasoning depth, matching effort to the task. Early testers, described by Mistral only as governments and enterprises, asked for more control over how long the model thinks, so the company added low and medium modes and made the high mode 50% more efficient.
• Higher coding scores. Terminal-Bench 4.0, which tests AI agents performing coding tasks in a terminal, rose from 28% to 41%.
• A quantised checkpoint. Mistral is also releasing a quantised version, meaning one stored at lower numerical precision to shrink memory needs, which runs on four Nvidia B200 or B300 GPUs.

For context, the preview was not cheap to run. At that stage, Artificial Analysis rated Large 4 roughly on par with OpenAI’s GPT-6 Luna and DeepSeek’s V4.1 Flash, but found it cost more than four times as much per task as open models of similar intelligence. The new token-efficiency claims are Mistral’s answer to that.

Large 4 specs: size, modalities and languages

Large 4 is a mixture-of-experts model: it has 1.05 trillion parameters in total but activates around 52 billion for any given task, which makes processing more efficient than running the whole network every time. It reads text and images and produces text and code.

Spec	Detail
Total parameters	1.05 trillion
Active parameters	About 52 billion
Inputs and outputs	Text and images in; text and code out
Reasoning modes	Low, medium, high
Languages	Trained on more than 160, including every official EU language
Context window	1M tokens, as listed by Shattered
Quantised version	Runs on four Nvidia B200 or B300 GPUs
Training hardware	3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centres

Sources differ slightly on training scale. Euronews and SiliconANGLE report 3,800 Grace Blackwell GPUs, while ZDNET quotes the company saying ML4 “was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months”. The reinforcement learning ran on Mistral’s own compute in France, drawing about 10 megawatts. The 1M-token context figure appears only in Shattered’s analysis, which also notes Mistral has not published long-context accuracy results.

Beyond coding, Mistral says the model can turn 2D blueprints into usable 3D models in CAD (computer-aided design) software, analyse large data volumes for financial decisions, and use satellite and aerial imagery to detect wildfires, track crop health and monitor risks to power lines.

How Large 4 compares with GLM-5.3 and other Chinese open models

The headline comparison is with GLM-5.3 from China’s Zhipu, which TechCrunch notes was previously considered one of the best open-weight models in the world. Almost every figure here comes from Mistral, and the outlets reporting them do not describe the result the same way.

• Terminal-Bench 4.0: TNW reports Large 4 at 41%, just ahead of GLM-5.3 at 40%. Fortune, citing Mistral’s figures, describes the two as tied.
• Artificial Analysis Coding Agent Index: Fortune reports Mistral’s claim that Large 4 narrowly tops all other open models on this index. ZDNET, however, reports that on the overall index, a composite of three benchmarks, Large 4 scores 48 to GLM-5.3’s 49, with OpenAI’s Codex GPT-6.1 Terra also at 48.
• Token efficiency: Mistral says Large 4 uses 47% fewer tokens per task on average than GLM-5.3, which TechCrunch notes makes it significantly more economical on the company’s account.
• Cyber defence: On CyberGym-E2E, a test of whether a model can find, reproduce and patch real memory-safety bugs in open-source C and C++ projects, Large 4 scored 79%, matching MiMo-V2.6-Pro and beating DeepSeek V4.1 Flash at 76%.

Closed models are still well ahead. Anthropic’s Mythos 5.1 scored 60.9% on the same Terminal-Bench 4.0 test, Fortune reports. ZDNET adds a general caution that benchmark scores should be taken with a grain of salt.

Stock also argued against Chinese open models on trust grounds. He told Fortune they may be subject to future export controls and that a model’s behaviour could be tweaked to exfiltrate secrets from a company’s systems. He told TNW that nothing guarantees Chinese labs will keep publishing weights. These are a competitor’s arguments, not findings about any specific model.

Access, licence and what it takes to host Large 4

The weights are on Hugging Face and the model is served through Mistral’s API. CNET lists further routes Mistral names: Mistral Studio, OpenRouter, Vibe Code and Nvidia-hosted endpoints. TNW describes Vibe access as limited to some users.

Licence terms are not described in the coverage. None of the reports reviewed sets out the conditions attached to the weights, such as limits on commercial hosting or revenue thresholds. Anyone planning to serve Large 4 to customers should read the licence on the Hugging Face model page before building on it.

Pricing is also open. API pricing for ML4 has not yet been published, according to TNW, and Stock argued that per-token rates mislead when a model needs far fewer tokens per job. Shattered likewise notes that any per-token figure circulating lacks a source.

Hardware is the practical floor. The quantised version needs four Nvidia B200 or B300 GPUs. Euronews points out that organisations self-hosting still need suitable hardware and expertise to operate the model securely. Shattered expects small teams to mostly use the API, with governments, large banks and telecom carriers more likely to justify their own clusters.

What Large 4 means for builders, and the risks

The main case for builders is control. Running the weights in-house lets teams keep sensitive material, such as proprietary code, away from outside AI services, and Mistral says self-hosters do not face a provider restricting or disabling the model. Stock framed it as business continuity, telling Fortune it is the top request from customers in finance, manufacturing and healthcare, who worry access to closed models “can be cut off”. Fortune points to June, when a US export-control directive forced Anthropic to suspend access to its Fable 5 and Mythos 5 models for nearly three weeks.

For agent workloads, the token claim matters more than a one-point benchmark gap. TNW notes that a 47% saving compounds when coding agents run for hours. Some customers already use Large 4 for automated penetration testing, running agents that probe their own applications and write up what they find.

The risks are real and acknowledged. Stock told UA.NEWS’s source that once released, access to the model cannot be fully restricted, which protects users from shutdown but also lets anyone use it. Mistral’s early testing partners got a less guardrailed version with “expanded cybersecurity capabilities”. Shattered argues that fine-tuning an open model can remove safeguards permanently once the files sit on an attacker’s disk, unlike jailbreaking a hosted model. Stock, asked about rogue AI agents, said: “We don’t want to fearmonger people. We think it’s a problem that is solvable”.

Mistral’s efficiency argument

Mistral frames Large 4 as proof that a smaller lab can reach the open frontier. CEO Arthur Mensch said:

“Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors. Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion.”

TNW notes that the claim of using a fraction of Chinese rivals’ resources is hard to test, given how little those labs disclose about training costs. Euronews adds that the reliance on Nvidia chips shows Europe still depends on foreign suppliers for key parts of the AI supply chain.

What to watch next

• Independent scores. Most figures are Mistral’s. Watch for Artificial Analysis and other evaluators to publish their own runs, especially to settle whether Large 4 ties or trails GLM-5.3 on the overall coding index.
• Pricing. Mistral has not published API rates.
• Smaller models. Distilled versions are due in the coming months, including one that runs on a laptop, without committed dates. Mistral also says more models from this new generation, spanning all modalities, will follow.
• Cadence. Stock says he wants Mistral not just to reach the frontier once but to “create the cadence of the frontier”.