Mistral AI Launches Codestral Coding GenAI Model
French artificial intelligence (AI) startup Mistral AI has made a significant stride in the tech world by introducing its new generative AI (GenAI) software, Codestral. This innovative model is designed to generate computer code, setting a new standard in the coding landscape.
Codestral: A Multilingual Code Maestro
Codestral is trained on a diverse dataset of over 80 programming languages, including popular ones such as Python, Java, C, C++, JavaScript, and Bash. This broad language base ensures that Codestral can assist developers in various coding environments and projects, making it a versatile tool in the coding community.
Enhancing Developer Productivity
One of the standout features of Codestral is its ability to complete coding functions, write tests, and finish partial code using a fill-in-the-middle mechanism. This capability not only saves developers time and effort but also helps them improve their coding skills by reducing the risk of errors and bugs.
Performance Benchmarks
Codestral is a 22B model that sets a new standard in the performance/latency space for code generation. It outperforms other models in RepoBench, a long-range evaluation for code generation, thanks to its larger context window of 32k. Evaluations across various benchmarks, including HumanEval, MBPP, CruxEval, and Spider, demonstrate Codestral’s superior performance in generating accurate and functional code.
Advanced API and Integration
Codestral operates through a shared instruction and completion API endpoint, which can be used to design advanced AI applications. Developers can interact with Codestral naturally and intuitively via Le Chat, Mistral AI’s free conversational interface, allowing them to leverage the model’s capabilities seamlessly.
Expanding in the U.S. Market
As part of its strategic expansion, Mistral AI has hired former Foursquare CFO Marjorie Janiewicz as its first U.S. general manager. This move is aimed at capitalizing on the growing demand from businesses seeking alternatives to AI models provided by major tech companies like OpenAI and Google. Mistral’s open-source approach, where the underlying code is openly shared and customizable, offers a safer and more versatile alternative to closed systems.
Community Feedback
The developer community has expressed positive feedback about Codestral. Nate Sesti, CTO and co-founder of Continue.dev, highlights that “a public autocomplete model with this combination of speed and quality hadn’t existed before, and it’s going to be a phase shift for developers everywhere.”
Similarly, Vladislav Tankov, Head of JetBrains AI, appreciates the strong focus on code and development assistance. Mikhail Evtikhiev, a researcher at JetBrains, noted impressive results with Codestral on their Kotlin-HumanEval benchmark, surpassing scores from GPT-4-Turbo and GPT-3.5-Turbo.
Getting Started with Codestral
Developers can download and test Codestral over HuggingFace, which is available under the Mistral AI Non-Production License for research and testing purposes. Codestral can be accessed via its dedicated endpoint, codestral.mistral.ai, and is also integrated into popular tools for developer productivity, including VSCode and JetBrains environments.
