Mistral AI made a groundbreaking announcement on Wednesday, July 24, 2024. They unveiled the highly anticipated Mistral Large 2, a revolutionary multilingual large language model (LLM).

This new version, aptly named Mistral Large 2, surpasses all expectations and far exceeds the performance of its predecessor, the February-released Mistral Large. In terms of mathematics, reasoning, code generation, instruction following, function calling, and support for a large number of languages, the new and enhanced model offers remarkable improvements over its predecessor.

The Mistral Research License, which permits unrestricted use and modification for academic and noncommercial purposes, was applied to the release of Mistral Large 2. While Mistral Large 2 availability for commercial deployment in IBM® watsonx is quite easy, business usage requiring self-installation necessitates contacting Mistral AI to get a Mistral Commercial License.

With no price increase, Mistral Large 2 has replaced Mistral Large in the watsonx Foundation model catalog, demonstrating IBM's dedication to giving customers access to the best and newest open models out there.

Transformer-based and dense, Mistral Large 2, also known as Mistral-Large-2407, is an LLM with 123 billion parameters. Unlike "sparse" expert architecture mixtures utilized by models such as Mistral's Mixtral-8x7B, "dense" in this context refers to standard neural network architecture.

In an LLM landscape where models typically go from 70 billion parameters to many hundreds of billions or even trillions of parameters, Mistral Large 2 carves out a special niche at 123 billion parameters. Mistral AI said that the size of Mistral Large 2 was intended to enable it to run at large on a single node in its official release announcement.

Despite the fact that the majority of cutting-edge LLMs (except for Meta's Llama 3 405B) are closed models with frequently undisclosed parameter counts, all of the information that is currently available suggests that Mistral Large 2 is competitive with top models despite having a much lower parameter count.