Mistral Unveils Large 2 to Rival Meta and OpenAI Models
Mistral introduces its new Large 2 AI model, claiming it matches or beats competitors in coding and math while using fewer parameters. The model also focuses on reducing hallucinations and offering broad multilingual support.
Mistral releases its new flagship AI model, Large 2, claiming it rivals the latest offerings from OpenAI and Meta in coding, math, and reasoning. The release comes just one day after Meta launches its massive Llama 3.1 405B model, highlighting a rapidly accelerating race in the frontier AI space. Mistral asserts that Large 2 outpaces Meta's new model on code generation and math benchmarks despite using fewer than a third of the parameters.
A major focus for Mistral during development is minimizing hallucinations, teaching Large 2 to simply acknowledge when it does not know an answer instead of fabricating a plausible response. The model also features a massive 128,000-token context window capable of processing roughly 300 pages of text at once. Additionally, Mistral highlights that Large 2 produces more concise outputs than competing models, which often suffer from unnecessarily verbose replies.
Despite its impressive performance, Mistral Large 2 lacks the multimodal capabilities that currently give OpenAI a significant edge in the industry. Furthermore, while Mistral markets the model as open, commercial applications require a paid license, and the computing infrastructure needed to run a 123-billion-parameter model remains prohibitively expensive for most organizations. The model is available now and includes broad multilingual support alongside proficiency in 80 coding languages.