Mistral AI is a newly established French artificial intelligence start-up. Mistral AI was established in 2023 (in the first half of the year, the specific month is unclear). Four weeks after its establishment, it received US$113 million in investment (seed round, valued at US$260 million). It is used to train and release large models and related products, which has attracted widespread global attention. This round of financing is also the largest financing received by a European large model start-up, and it is also regarded as the light of European large model companies~

Some (perhaps most) members of Mistral AI come from DeepMind and Meta companies and were engaged in large model-related work before joining MistralAI. The main goal of MistralAI is to provide generative AI technology to the world to help solve human problems.

On September 27, 2023, MistralAI released their first large language model Mistral 7B (model information card address: https://www.datalearner.com/ai-models/pretrained-models/Mistral-7B), claiming to be the best large model with a scale of 7 billion parameters. The parameter of Mistral 7B is 7.3 billion, which exceeds the LLaMA2-13B model in all evaluation benchmarks. At the same time, its code capability is close to CodeLlama 7B, and its performance is very impressive. The most important thing is that this model is open source under the Apache 2.0 license, which means it is very friendly for commercial use.

On December 8, 2023, the company announced a download link for an 8×7B Mixture of Experts model on its official Twitter account, which is almost the best current open source implementation of MoE (MoE is one of the very important optimization technologies in the field of large models, specific reference: https://www.datalearner.com/llm-blogs/deep-dive-into-mixture-of-experts-in-large-models). This method of releasing a large model download link without introducing anything has also caused a lot of discussion in the community.

The biggest feature of Mistral AI is that the products currently released are very transparent and open. Although it only released a large language model in September 2023, its friendly open source agreement and transparent technology sharing have been praised by many people. On the official website of Mistral AI, they introduced their concept. Mistral AI believes that although previous private model services (probably mainly ChatGPT) have proven the capabilities of generative large models, historically noted technologies or products such as browsers (Webkkit), operating systems (Linux), cloud computing (Kubernets), etc. all have excellent open source community leaders, and open source solutions will quickly surpass private solutions in most use cases. Mistral AI firmly believes that this path will also help large models overcome problems such as bias. Therefore, MistralAI chooses to use open source to release its products.

Currently, MistralAI has released many models and products, including:

Model name Model introduction DataLearnerAI information card address of the model

Mistral-7B is one of the most powerful large language models with a scale of 7 billion parameters. It has excellent instruction compliance and multi-language support. https://www.datalearner.com/ai-models/pretrained-models/Mistral-7B-Instruct

Mixtral-8×7b, the most famous mixed expert large model, directly releases pre-training results, which has attracted widespread attention. The effect exceeds GPT-3.5. https://www.datalearner.com/ai-models/pretrained-models/Mistral-7B-MoE

Mistral Medium closed source model, performance is close to GPT-4 in all aspects, but the price is cheaper

The mysterious model released by Mistral NextMistralAI appeared on the Areno experience on February 16, 2024, without any information, and the evaluation results are close to GPT-4. https://www.datalearner.com/ai-models/pretrained-models/mistral-next