Coding with AI

Mistral’s Le Chonk has emerged as a top open-source coding model, ranking alongside Codex GPT-6.1 Terra.

Mistral released ML4 weights, Le Chonk, ranking near top open coding models with lower token use and stronger cyber defense.

Mistral’s Le Chonk has emerged as a top open-source coding model, ranking alongside Codex GPT-6.1 Terra.

Key Takeaways

  • Mistral released open weights for Mistral Large 4, nicknamed “Le Chonk,” giving the public full access after its Oct. 6 preview.
  • ML4 ranks among the top open-source coding models, scoring 48 on the Artificial Analysis Coding Agent Index and using 47% fewer tokens per task than comparable open models.
  • The model emphasizes cyber defense capabilities, targeting enterprise and government security needs while offering user-controlled, sovereign deployment.
  • ML4 was trained from scratch on 4,000 Nvidia GPUs in Europe, positioning it as a smaller-scale Western open model competing with larger proprietary and Chinese rivals.
  • While benchmark results are close and should be treated cautiously, the release signals growing competitiveness of Western open models in coding and AI security.

Meta: Mistral releases ML4 weights after Oct. 6 preview, with Le Chonk scoring 48 on the Artificial Analysis Coding Agent Index and using 47% fewer tokens.

Mistral’s Le Chonk has emerged as a top open-source coding model, ranking alongside Codex GPT-6.1 Terra.

Featured image: Abstract, friendly illustration of model weights flowing into a security shield, with no real people, logos or text.

Tags: Mistral, ML4, open weights, coding benchmarks, AI cybersecurity

For developers, enterprises and government teams comparing open AI models, Mistral has released the weights for Mistral Large 4 (ML4), nicknamed Le Chonk for its trillion-parameter size, after previewing it on Oct. 6. The French AI lab says the release gives public access to the model, not just preview access, and its new benchmark scores show Western open models gaining ground over Chinese peers.

Mistral says this sets a new standard for Western-made open models. After three weeks of testing with a select group of partners, ML4 ranks at the top of Artificial Analysis’ Coding Agent Index among open models globally. It also now offers three reasoning modes, low, medium and high, a pattern common among larger models that lets users trade effort against task complexity.

A higher open-model coding score with fewer tokens

On Terminal-Bench 4.0, an agentic coding benchmark, ML4 improved from 28% in preview to 41%, placing it one small but important point ahead of Opencode’s GLM-5.3. The broader Artificial Analysis Index, which shows a composite score across three relevant benchmarks, still favors GLM-5.3: ML4 scores 48, while GLM-5.3 scores 49. Codex GPT-6.1 Terra also scores 48 overall.

Mistral frames the result as an efficiency story. The company said ML4 “rivals top open models such as GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% less tokens per task on average.” That claim matters because proprietary AI remains expensive, and users often need to decide whether a model’s score comes with a heavier compute or token bill.

Mistral’s Le Chonk has emerged as a top open-source coding model, ranking alongside Codex GPT-6.1 Terra.

The gains arrived after only a few weeks of extra training, but benchmark numbers deserve caution, especially because many models can cheat. For readers, the practical takeaway is not that one point decides a model, but that an open Western model now sits close to the top open coding tier while promising lower token use.

Cyber defense becomes the second battleground

Mistral used its Oct. 6 preview to emphasize cyber defense capabilities, customizable control and data sovereignty. Over the following weeks, it worked with developers and leaders to assess ML4’s cybersecurity behavior. That process mirrors a broader trend among proprietary American labs such as OpenAI, Google and Anthropic as security concerns mount.

As with Anthropic’s Project Glasswing and OpenAI’s rollout of Astra, Mistral gave initial testing partners access to a less guardrailed version of ML4 with expanded cybersecurity capabilities. On the CyberGym benchmark for defensive cybersecurity, Le Chonk scored just above DeepSeek V4.1 Flash. CyberGym measures how well models find real memory-safety vulnerabilities in C/C++ open-source projects, reproduce them with proof-of-concept inputs and patch them without breaking the projects’ tests.

The release lands in a security debate that has pitted open and proprietary models against each other. Open models were first described as riskier because they can be modified, but recent AI security incidents have led some defenders to view open weights as a useful alternative after proprietary models from Anthropic and OpenAI were shown to carry risk too.

Mistral says ML4 is built for security that stays under user control, unlike proprietary models whose access frontier labs can technically rescind at any time. “The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks,” Mistral co-founder Guillaume Lample said during a briefing for the Oct. 6 preview release.

Mistral’s Le Chonk has emerged as a top open-source coding model, ranking alongside Codex GPT-6.1 Terra.

Following the Hugging Face breach, Mistral joined many companies signing Nvidia’s Open Secure AI Alliance, a cross-industry partnership that argues open models are crucial to democratizing defenses against increasingly common AI security incidents. By Nvidia’s logic, the race is between Mistral and other open models to reach state-of-the-art security prowess while keeping defenses broadly available.

Training on 4,000 GPUs in Europe

Training a competitive model during a compute shortage is difficult for a smaller lab like Mistral, which does not have the same resources as pre-IPO giant Anthropic.

The company said ML4 “was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, deployed in Mistral’s own data centers in Europe.” For context, Nvidia CEO Jensen Huang said on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs.

That comparison frames ML4 as a smaller, more targeted build. Its claim to compete on coding and defensive cybersecurity while using fewer tokens per task is the core pitch: a Western open model that tries to deliver more capability from less compute.

What the weights release changes

The Oct. 6 preview introduced Le Chonk; the weights release moves ML4 from a tested preview into broader public access, giving users more direct control over the model. The benchmark table still shows GLM-5.3 ahead by one point overall, and Codex GPT-6.1 Terra level with ML4 at 48, so the open-model race remains close.

For readers, the value is in the combination: top open coding ranking, a small but visible efficiency claim, cybersecurity focus and European training infrastructure. The next test is how much of that performance holds up once the weights are in the hands of many users, especially in security-sensitive coding and defensive tasks.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.