Mistral, an AI startup based in Paris, secured the largest-ever seed funding in Europe one year ago, positioning itself as a major player in the global AI race. The investment highlights the confidence investors have in Mistral and points to the increasing strategic importance of advanced AI technologies in Europe’s tech ecosystem.

Mistral recently expanded into the programming and development sector with the introduction of Codestral, a code-centric large language model (LLM).

Codestral stands out as it is available under a non-commercial license, offering developers and researchers a powerful tool without the immediate pressure of commercial application. This move aims to foster innovation and allow the tech community to explore and expand the capabilities of AI in coding.

Technical specifications and language support of Codestral

Core features and specifications

Codestral is a cutting-edge AI model with 22 billion parameters, classified as an open-weight generative model. It’s engineered to handle a full range of coding tasks, from initial generation to the final completion stages.

One of the standout features of Codestral is its substantial context length of 32K, which aids developers working across various coding environments. Extensive context length is key in understanding longer blocks of code, enabling more accurate and context-aware code generation.

Training and supported programming languages

Mistral has trained Codestral on a diverse dataset that includes more than 80 programming languages, making it exceptionally versatile. The model’s ability to generate code, complete partial functions, and conduct robust testing makes it a comprehensive tool for developers.

Codestral supports both widely-used programming languages such as SQL, Python, Java, C, and C++ and more niche languages like Swift and Fortran. Broad language support makes sure that Codestral can meet the needs of a diverse developer base, catering to both mainstream and specialized programming requirements.

Performance metrics and developer benefits

Boosting developer productivity

Codestral is designed to elevate developer productivity by optimizing workflows. With its advanced capabilities, Codestral reduces the time and effort developers invest in building applications.

The model assists in minimizing coding errors and bugs, which are often costly and time-consuming to correct. Proactive work on error reduction is critical in software development, where the early detection and resolution of issues can greatly expedite product timelines and improve software reliability.

Comparative performance

Codestral positions itself as a superior option compared to other prominent code-centric models like CodeLlama 70B, Deepseek Coder 33B, and Llama 3 70B. Mistral’s strategic focus on developing a model that excels across a broad range of programming tasks highlights its push to lead the industry in the generative coding space.

Here’s an overview of Codestral’s performance statistics:

  • RepoBench performance: Codestral shows its robust capabilities in Python code completion with an accuracy of 34%. This benchmark is particularly challenging, and Codestral’s performance was notable.
  • HumanEval and CruxEval scores: Codestral scores impressively in the HumanEval test with an 81.1% success rate for Python code generation. It also achieves a 51.3% on CruxEval for Python output prediction. These scores are important as they show the model’s ability to understand and predict code behavior accurately, which is key for developing functional and reliable software applications.
  • Multi-Language proficiency: In a comprehensive assessment involving multiple programming languages such as Bash, Java, PHP, C++, C, and TypeScript, Codestral leads with an average score of 61.5%. Versatility is a priority for developers working in multi-language environments and makes sure that Codestral is applicable across various software development scenarios.
  • SQL performance on Spider: Codestral secures the second position in the Spider assessment for SQL performance with a score of 63.5%. This performance is indicative of its strong capabilities in database scripting, a key area of software development that demands precision and efficiency.

Accessibility, licensing, and initial market adoption

Availability and licensing options

Mistral offers Codestral on Hugging Face with a non-production license that restricts use to non-commercial purposes. Strategic licensing allows developers, researchers, and enthusiasts to explore Codestral’s capabilities without a financial commitment, fostering a broad user base from the outset.

Mistral provides access to Codestral through two distinct API endpoints:

  • codestral.mistral.ai: Designed specifically for use within Integrated Development Environments (IDEs), this endpoint caters to developers seeking to integrate Codestral’s capabilities directly into their workflow. It uniquely features a personal API key management system, free from typical organizational rate limits.
  • api.mistral.ai: This endpoint targets a wider range of uses including broader research activities and third-party application development, with usage billed per token. This flexibility makes it suitable for larger-scale projects that may require extensive API calls.

During an initial eight-week beta period, the IDE-specific endpoint is available at no cost, providing a significant incentive for early adopters to integrate and test Codestral in real-world scenarios.

Developer tools and feedback

Several prominent developer tools and platforms are currently testing Codestral. These include LlamaIndex, LangChain, Continue.dev, Tabnine, and JetBrains—each a leader in the software development and AI sectors. The testing by these entities validates the potential of Codestral and accelerates its integration into mainstream development processes.

Early feedback from these tests highlights Codestral’s rapid performance and its expansive context window as particularly beneficial.

For instance, the successful use of Codestral in self-corrective code generation, as tested with LangGraph, showcases its practical utility and immediate applicability in enhancing code quality and efficiency.

Competitive landscape and future outlook

Codestral enters a highly competitive market, where it faces off against established models such as StarCoder2, OpenAI’s Codex and GPT-4 Turbo, Amazon’s CodeWhisper, and smaller but agile models from Replit. Each competitor brings distinct capabilities and market approaches, challenging Mistral to continuously innovate and improve Codestral.

Codenium, another major competitor, has recently secured $65 million in Series B funding, achieving a valuation of $500 million. This investment spotlights robust investor confidence in AI-driven coding solutions and highlights the lucrative nature of this market segment.

In this dynamic and rapidly advancing field, Codestral’s success will depend on its technological prowess and on Mistral’s ability to adapt to market needs, respond to developer feedback, and iterate its offerings effectively.

Strategic decisions around accessibility and licensing are likely to further position Codestral as a primary contender in the race to dominate AI-augmented coding tools, promising an exciting future as the technology evolves.

Tim Boesen

June 10, 2024

5 Min