On October 29, Mistral made its Large 4 weights available for download. This means that businesses and governments can implement the cybersecurity AI model in-house rather than depending on American or Chinese companies. However, this freedom comes at a price. Since Mistral has little control over the model, it cannot determine who will be able to use it and for what purposes.
From API preview to downloadable weights
Mistral’s Large 4 was first rolled out via an API on October 6 and its weights were issued for download on October 29. API access is still possible as a deployment option. The official specifications of Large 4 indicate its parameters are 1.05 trillion and include the active parameters being 52 billion, which is marginally more than the 49 billion as covered earlier by Cryptopolitan .
Businesses can customize and host the model themselves. But those seeking to resell it as a hosted service need a separate commercial license. That independence is central to Mistral’s pitch. As company executive Pierre Stock told TechCrunch , access “can’t be cut off.”
What the benchmarks show, and what they do not
Mistral reports that Large 4 scored 48% on the Coding Agent Index, putting it nearly level with GLM-5.3. However, Large 4 used 47% fewer tokens per task. The company also reported scores of 19.7% on FrontierCode 2.0 and 41% on Terminal-Bench 4.0.
The picture is slightly different for cybersecurity. Large 4 scored 51 on the Artificial Analysis Cyber Index , trailing Kimi K3.1 Pro (64), DeepSeek V4.1 Flash (57), and Grok 4.7 (56). But these scores need context. The index measures how well AI models find and fix security vulnerabilities. It does not assess their ability to carry out cyberattacks.
Mistral Large 4 Cyber AI Benchmarks vs Rivals and Open-Weight Adoption
How far behind the closed frontier
Despite Large 4’s reported improvements in coding and defensive cybersecurity, leading open-weight models still trail the strongest closed systems. An October 16 Epoch AI analysis gave GLM-5.3, the leading open-weight model, a Cyber ECI score of 160. The strongest closed models scored 171. Looking at historical performance, Epoch estimated that open-weight models were six to seven months behind the closed frontier. However, its assessment came before Large 4’s final release.
Britain’s AI Security Institute reached a similar conclusion. It estimated a four-to-seven-month gap of open-weight models from its closed-weight rivals, although its tests differed from Epoch AI’s.
The issue extends beyond the matter of defensive capabilities. As part of its controlled tests, Anthropic reported that GLM-5.3 successfully exploited existing vulnerabilities in 50 out of 410 trials. Researchers managed to get around the safety features of the model in 64%-100% of their various simulations.
The sovereignty pitch meets adoption reality
Companies are showing more interest in controlling their AI systems, but spending tells a different story. According to reports, 42% of surveyed enterprises used open-weight models in production in September, up from 31% in July.
Still, open weights accounted for just 5% of metered AI consumption in October. Claude held 29%, while Microsoft 365 Copilot accounted for 36%.
Futurum also found that model origin influenced decisions for 47% of production users. However, its survey focused largely on North America, limiting broader conclusions.
For businesses, greater independence also means taking on more responsibility. PwC has warned that companies choosing open weights must manage their own security and maintenance.
Safeguards that cannot be recalled
That responsibility becomes especially important when AI models can handle sensitive cybersecurity tasks. Mistral says Large 4 refuses 96.6% of harmful cyber requests. However, those safeguards can be removed once the model is downloaded. AISI also warns that developers cannot monitor or withdraw privately hosted copies.
Moonshot has taken a different approach. The company removed cybersecurity training material from Kimi K3.1 Pro before its planned November 4 weights release. Its internal offensive benchmark score subsequently fell from 90.5% to 81.0%. However, those results have not been independently verified.
As of October 30, no publicly documented cyberattack has been linked to Large 4’s newly released weights.
Don’t just read crypto news. Understand it. Subscribe to our newsletter. It’s free .
