Mistral Large 2 Delivers Frontier Performance With Reduced Hallucinations
Mistral AI releases Mistral Large 2, a 123-billion-parameter model that rivals top competitors in coding and reasoning while significantly reducing hallucinations. The model features a massive 128k context window and supports dozens of human and programming languages.
Mistral AI unveils Mistral Large 2, a powerful new 123-billion-parameter language model that pushes the boundaries of speed, performance, and cost efficiency. The model features a 128k context window and supports dozens of human languages, including French, German, Spanish, Arabic, and Chinese, alongside over 80 coding languages like Python, Java, and C++. Its size allows it to run at high throughput on a single node, making it highly practical for long-context applications.
The new model excels significantly in code generation and reasoning tasks, performing on par with leading industry models such as GPT-4o and Claude 3 Opus. Mistral achieves this by training the model on a very large proportion of code, building directly on their previous experiences with the Codestral family. Furthermore, Mistral Large 2 sets a new standard on the performance versus cost Pareto front for open models, achieving an 84.0% accuracy on the MMLU benchmark in its pretrained form.
A major focus during the development of Mistral Large 2 is the drastic reduction of hallucinations. Developers fine-tune the model to be more cautious and discerning, ensuring it provides reliable, factual outputs instead of plausible-sounding but incorrect information. Additionally, the system is trained to explicitly acknowledge when it lacks sufficient information to provide a confident answer, a feature that directly contributes to its improved performance on mathematical and problem-solving benchmarks.