Mistral Lays Claim to AI Coding Crown

Mistral today made accessible the weights for the Mistral Large (ML4) model that have now been tuned to become what the company claims now provides the agentic coding capabilities of any model available.

Pierre Stock, vice president of science at Mistral, said that after three weeks of reinforcement learning and testing since first launching the 1-trillion-parameter ML4 model, also known as Le Chonk, the coding capabilities of the multimodal model have significantly improved after adding two additional reasoning modes: Low, medium, and high, that enable 52 billion parameters to be active at any given time.

As a result, ML4 now tops open models such as GLM-5.3 from Z.ai on the Artificial Analysis Coding Agent Index while using 47% less tokens per task on average. On the Terminal-Bench 4.0 benchmark, which tests agentic coding in the terminal, ML4 climbed from 28% to 41%, tying GLM-5.3 as the best model for coding agents.

Additionally, ML4 scores highly in terms of its ability to discover vulnerabilities in code at a much lower cost than rival frontier models, noted Stock. The challenge, of course, is that cybercriminals may also be able to access those same models to now develop exploits in a few hours for any vulnerability discovered

After only raising a comparatively smaller $6 billion to build its models, Mistral is demonstrating that open-weight models have become a viable alternative for a wide range of use cases. Each organization will need to determine to what degree to rely on open-weight models, but as the cost of AI coding continues to rise, many DevOps teams are exploring their options. Up until now, the best open-weight alternatives were all developed in China, which for many organizations raises security concerns. Mistral is making a case for an open-weight alternative that is primarily developed in France.

DevOps teams can then opt to deploy Mistral models on IT infrastructure they manage directly or rely on any number of cloud service providers that host those models.

Of course, the performance levels that any DevOps team is likely to achieve is going to differ from what a model provider achieves using a set of benchmarks, but in most cases DevOps teams don’t always need access to the latest models that have been developed. Many lower-cost models are sufficient for the purpose of generating code and even discovering vulnerabilities in legacy code.

Ultimately, most DevOps teams will wind up employing multiple AI models. The challenge is determining which model to use when for a specific task. In many cases, DevOps teams are now integrating routers and gateways into the harnesses they have deployed to invoke a specific set of AI models.

It’s not clear how long open-weight models will be able to lay claim to providing the best performance. Providers of AI models are constantly leapfrogging one another in terms of the capabilities they are able to provide. Regardless of which AI models are deployed, however, the cost of the tokens used to create code is becoming a more significant DevOps concern as the tasks assigned to AI models and agents become more complex with each passing day.