French AI leader Mistral has unveiled its new flagship model, mistral-large-2407, achieving a stunning 84% on the MMLU benchmark and placing it in direct competition with top models like GPT-4o and Claude 3 Opus. Announced at the company's AI Now Summit in Paris, the release also includes a new open-weight coding model and major upgrades to its API platform, La Plateforme. This move solidifies Mistral's position as a top-tier global player in the generative AI race.
A New Flagship to Challenge the Titans
The star of the show, mistral-large-2407, is the successor to Mistral's previous top model and represents a significant leap in capability. According to detailed notes from the summit published by developer Koen van Gilst, its 84% score on the Massive Multitask Language Understanding (MMLU) benchmark puts it squarely in the performance tier of industry leaders.
Mistral has armed its new model with a suite of developer-focused features designed for practical application, moving beyond raw benchmark performance.
Key enhancements for mistral-large-2407 include:
- Native Function Calling: Enables more reliable and structured interactions with external tools and APIs.
- Native JSON Mode: Guarantees valid JSON output, a critical feature for developers building applications.
- 128k Context Window: Allows for processing and reasoning over large documents and complex prompts.
- New Embedding Model: A new
mistral-embed-2model is optimized for Retrieval-Augmented Generation (RAG) tasks.
Codestral-22B: A New Force in AI-Powered Coding
Alongside its flagship model, Mistral released codestral-22b-instruct, a powerful 22-billion parameter model specialized for coding tasks. Trained on over 80 programming languages, it aims to be a versatile assistant for developers.
Codestral-22B reportedly outperforms larger models like CodeLlama 70B on key benchmarks such as HumanEval and CruxEval. It also introduces a "windowing" technique that extends its native 32k context window to an effective 128k, a first for a model of its size.
Platform Upgrades and a Shift in Open Strategy
Mistral is also investing heavily in its developer platform, La Plateforme. The company announced "Le RAG," a managed RAG-as-a-service solution that simplifies the process of building context-aware AI applications by handling data indexing, chunking, and retrieval automatically.
The company addressed its commitment to open source, clarifying that while it will continue to release open-weight models, its most advanced, "frontier" models like will remain proprietary. For ongoing analysis of these evolving AI business models, be sure to subscribe to the AI Breaking Wire newsletter for weekly insights delivered to your inbox.