Artificial intelligence startup Mistral AI today debuted Codestral, a large language model optimized for software development tasks.

The launch comes amid reports that the company is in the process of raising a sizable new funding round. Earlier this month, sources told Wall Street Journal that Mistral is seeking $600 million from investors at a $6 billion valuation. That’s triple what Mistral was worth following its most recent funding round in December.

The company’s new Codestral model understands more than 80 programming languages. Its knowledge base includes so-called high-level languages such as Python that automate certain coding tasks to improve developer productivity. The model can also write software in several low-level syntaxes, which enable programmers to more directly interact with the underlying hardware. That feature is conducive to tasks such as optimizing programs’ performance but comes with a steep learning curve.

Mistral says that Codestral lends itself to a range of coding tasks. Developers can upload a snippet of code and have the model explain what it does. Additionally, Codestral is capable of generating new code based on natural language instructions.

The model can function as a kind of autocomplete tool and continue a snippet of code that a developer has started writing. It’s also capable of modifying components of an application that are already complete. A developer could, for example, ask Codestral to change the middle line of a code snippet written by a colleague.

Another task that the LLM promises to ease is the process of testing newly created software for bugs. 

Some bugs prevent an application from loading, while others only emerge when users perform a specific sequence of actions in the program’s interface. The latter type of issue can be highly challenging to detect using manual troubleshooting methods. According to Mistral, developers can use Codestral to automatically scan their code for flaws and thereby speed up the process. 

Codestral features 22 billion parameters. In an internal evaluation, Mistral compared it against three other open-source LLMs including Meta Platforms Inc.’s Llama 3 70B, which includes more than three times as many parameters. Codestral proved more adapt than all three models at Python programming tasks and achieved the second-highest score in a test that assessed their SQL prowess.

One contributor to the LLM’s performance is its large context window. Codestral can process prompts with up to 32,000 tokens, more than twice the amount of input data supported by Llama 3 70B. That makes it easier for the former LLM to analyze large code files that contain up to hundreds of lines of application logic.

Codestral is available under an open-source license for research and testing purposes. Organizations that wish to use the model in commercial software projects can access it through a cloud-based application programming interface provided by Mistral. According to the company, the API lends itself to tasks such as building programming automation plugins for code editing applications. 

Image: Unsplash