Mistral unveils new AI model Large 2, takes on Meta’s Llama 3.1 and OpenAI’s GPT-4o

It said that with Mistral Large 2, the company continues to push the boundaries of cost efficiency, speed, and performance. The new AI model is available on its la Plateforme, Mistral’s platform that offers access to the company’s large language models (LLMs). The company says the new model comes with new features that allow developers to build innovative AI applications.

What is Mistral Large 2?

Mistral Large 2 is an AI model from Mistral, essentially a predecessor to Mistral Large 1 which was launched in February and was out of all the AI models, it was reportedly only second to OpenAI’s GPT-4.

The new Large 2 comes with a 128k context window and supports over a dozen languages including Hindi, Russian, French, Spanish, German, Italian, Arabic, Portuguese, Chinese, Korean and Japan. The AI model also supports over 80 coding languages including C, C++, Java, JavasScript, Python, Bash, etc. When it comes to parameters, Large 2 comes with 123 billion parameters, much less than Meta’ Llama 3.1 405B. Regardless, the company claims that Large 2 outperformed Llama 3.1 in math and code generation.

With Large 2, Mistral claims that it has minimised hallucinations and produces more concise responses than leading AI models. The model is currently available on Le Chat, Mistral chatbot similar to ChatGPT. To access Le Chat, one simply needs to sign up with a valid email address.

How powerful is Mistral Large 2?

Mistral Large 2 not only excels in performance but also cost efficiency, the company claims that it achieved an 84.0 per cent accuracy on MMLU – a new standard for open models. It has been reportedly trained extensively on code, and surpasses its predecessors significantly and even matches top models like Anthropic’s Claude 3 Opus, GPT-4o, and Llama 3.1 405B. Mistral said that with the new model, it has focused on enhancing its reasoning and considerably reducing hallucinations or false information.

The company further said that the model has been made more cautious, generating reliable outputs. Large 2 has even been trained to admit when it doesn’t have the right information, and this has led to it performing better on mathematical benchmarks and has enhanced its problem solving skills.

According to the company, Large 2 is a high-performing model especially when it comes to coding. It achieved 76.9 per cent accuracy across various programming languages on average. When compared to Large 1, which has an average 60.4 per cent accuracy, Large 2 displayed significant improvement. Furthermore, it outperforms models like Llama 3.1 405B and Llama 3.1 70B in several languages and is on par with GPT-4o in many benchmarks, Mistral said.

Based on the company’s data, Large 2 is evidently strong in Python and TypeScript. This makes it ideal for developers working with these languages. It also showed good performance in Java and PHP, the advancements show Large 2 is capable of handling complex programming tasks with higher accuracy. It is important to note that the benchmark scores shared by the companies usually don’t reflect the claims as most companies cherry-pick scores making it difficult for one to get a full picture of the model’s capabilities.