Skip to main content

Kaer NewsSoftware

Mistral releases open weights for trillion-parameter Large 4 model

French AI lab Mistral released the weights for its trillion-parameter Mistral Large 4 model on 29 October 2026, letting businesses and governments download and run the system themselves.

4 min read
Mistral releases open weights for trillion-parameter Large 4 modelPhoto: TechCrunch

Key points

Mistral released the open weights for its trillion-parameter Mistral Large 4 model, letting enterprises and governments download and deploy it themselves.

French AI lab Mistral released the open weights for its trillion-parameter model Mistral Large 4, nicknamed Le Chonk, on 29 October 2026, letting businesses and governments download and deploy the system themselves. Weights are the trained numerical values that make a model work. The release lets organisations run ML4 on their own infrastructure rather than calling a remote API.

The release matters because open weights let enterprises and governments keep control of their data and systems, an argument Mistral says carries special weight in cybersecurity, where outside providers could cut off access. Mistral CEO Arthur Mensch said AI should "be a tool for empowerment, not dominance." The company claims state-of-the-art performance at a fraction of the compute and capital committed by closed labs.

What the release changes

Mistral previewed ML4 on 6 October 2026 and released the weights three weeks later, on 29 October 2026. The weights are available on Hugging Face and via Mistral's API. Pierre Stock, Mistral's Vice President of Science, said open weights mean access "can't be cut off by definition because the model is open-weighted," letting firms deploy it on-premises in their own ecosystem.

Mistral says ML4 matches Chinese open-weight rival GLM-5.3, previously one of the best open-weight models, particularly on cyber work, while using 47% fewer tokens per task — tokens are the text units models process, so fewer tokens per task means lower cost. On Artificial Analysis' Coding Agent Index, Mistral claims ML4 ranks first among open models globally for agentic coding.

On Terminal-Bench 4.0, an agentic coding benchmark, ML4 scored 41%, up from 28% in preview, one point above Opencode's GLM-5.3. On the overall Index, a composite of three benchmarks, ML4 scores 48 against GLM-5.3's 49; Codex GPT-6.1 Terra also scores 48. On the CyberGym defensive cybersecurity benchmark, ML4 scored just above DeepSeek V4.1 Flash.

Where the benchmark numbers came from

ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months in Mistral's own European data centres. For comparison, Nvidia CEO Jensen Huang said on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs — about 25 times as many. ML4 offers three reasoning modes: low, medium and high, matched to task complexity.

Mistral frames open weights as sovereign cyber defence: companies can download, deploy and own the model to detect, analyse and respond to threats across code and IT infrastructure. Stock said organisations gain frontier intelligence to defend against cyberattacks, with access that cannot be cut off. Guillaume Lample, Mistral co-founder, said the capabilities target threat actors jailbreaking closed models to perform cyberattacks.

The dual-use trade-off

The trade-off is that freely downloadable models can also be customised by attackers. Rafe Pilling, senior director of threat intelligence at security company Sophos, called open-weight models "clearly a dual-use technology" that can be abused as well as used productively. Stock said Mistral wants to "provide this same capability to everyone as soon as possible" because some malicious actors might use the technology to attack.

Stock also addressed rogue AI-agent incidents, saying "We don't want to fearmonger people. We think it's a problem that is solvable." Pilling argued all advanced models can go rogue without careful supervision regardless of origin, saying the technical realities and the nationalistic narrative are two different things, and what matters is having the right monitoring and controls to detect deceptive behaviour.

Mistral's testing partners received a less guardrailed version of ML4 with expanded cybersecurity capabilities, similar to Anthropic's Project Glasswing and OpenAI's rollout of Astra. After the Hugging Face breach, Mistral signed Nvidia's Open Secure AI Alliance, a cross-industry partnership arguing open models are crucial to democratising defences against AI security incidents.

Mistral claims ML4 rivals top open models such as GLM-5.3 while using 47% fewer tokens per task on average, making it cheaper to run at the same performance level. The company also claims ML4 outranks all other open models on Artificial Analysis' Coding Agent Index for agentic coding tasks, though benchmark scores should be treated with caution given that some models cheat.

The weights are live on Hugging Face and Mistral's API as of 29 October 2026, so enterprises and governments can begin downloading and deploying ML4 immediately. No further dated milestones for the release appear on the pages reviewed; the next step is adoption by organisations seeking on-premises, sovereign cyber defence capability that cannot be switched off remotely.

Frequently asked questions

What did Mistral release on 29 October 2026?

Mistral released the open weights for Mistral Large 4 (ML4), its trillion-parameter model nicknamed Le Chonk, letting businesses and governments download and deploy it themselves via Hugging Face or Mistral's API.

Why does Mistral say open weights matter for cybersecurity?

Mistral says open weights let enterprises and governments run ML4 on-premises with access that cannot be cut off remotely, unlike proprietary models from OpenAI, Anthropic or Google that providers could revoke.

More from this edition