Mistral Unveils Large 2 Model to Rival Meta and OpenAI
French AI startup Mistral launches its new Large 2 model, claiming it matches or beats recent releases from Meta and OpenAI in coding and math while using fewer parameters. The 123-billion-parameter model focuses on reducing hallucinations and offers a massive 128,000-token context window.
Mistral releases its new flagship AI model, Large 2, positioning it as a direct competitor to the latest offerings from OpenAI and Meta. The Paris-based startup claims this model matches or exceeds the performance of Meta's recently released Llama 3.1 405B in code generation, mathematics, and reasoning. Remarkably, Mistral achieves these results with only 123 billion parameters, which is less than a third of Meta's massive model.
A major focus during the development of Large 2 is the reduction of hallucinations, a common flaw in generative AI. Mistral trains the model to recognize its own limitations and admit when it does not know an answer rather than fabricating a plausible response. Additionally, the model features a 128,000-token context window, allowing it to process roughly 300 pages of text in a single prompt, and supports a wide array of languages alongside 80 coding languages.
Despite its impressive capabilities, Mistral Large 2 lacks the multimodal features that currently give OpenAI a significant edge in the AI race. Furthermore, while Mistral markets the model as open, commercial applications require a paid license, and the computing infrastructure needed to run a model of this size remains out of reach for most organizations. This release follows Mistral's recent $640 million Series B funding round, which values the fast-growing startup at $6 billion.