Paris-based startup Mistral AI is stepping up into the big leagues, launching Mistral Large to compete with other top-tier large-language models and unveiling a beta version of “Le Chat,” its consumer-facing chatbot intended to rival market leader Open AI’s ChatGPT.
“Mistral Large is our flagship model, with top-tier reasoning capacities,” the company said in an official announcement, “Mistral Large achieves strong results on commonly used benchmarks, making it the world's second-ranked model generally available through an API (next to GPT-4).”
Mistral Large supports a context window of 32K tokens, generally more than 20,000 words in English, and is fluent in English, French, Spanish, German, and Italian with a nuanced grasp of grammar and cultural context for each. The startup says its flagship model “is ideal for complex tasks that require large reasoning capabilities or are highly specialized” and describe its outputs as “concise, useful, unopinionated, with fully modular moderation control.”
Mistral AI has been a darling of the open-source AI community thanks to its high-performing open-source models like Mistral 7B and the top-of-the-line Mixtral 8x7B, which used a mixture-of-experts approach to increase its overall quality. However, Mistral Large is a proprietary model, so there is limited technical information to independently compare this model against its competitors.
The company didn’t respond to a request from Decrypt for related technical papers or details about the number of training parameters, training methods or even the data corpus used to build the model.
How does Mistral Large stack up against its competitors, at least based on its creator's tests?
Mistral AI claims that Mistral Large ranks second after GPT-4 based on several benchmarks, but real-life usage may always vary. Mistral Large has not been tested in third-party rankings like the Chatbot Arena, but Mistral AI claims it would outperform Mistral Medium, which has ranked better than GPT-3.5, Claude1, Claude 2, and Qwen based on blind comparisons of outputs with similar prompts.
Mistral Large is now available through a paid API, and is a lot cheaper than OpenAI’s option. Mistral Large costs $8 per million tokens of input and $24 per million tokens of output (the same as Claude), whereas GPT-4 costs $10 and $30, respectively.
Le Chat, Mistral AI's chat assistant, is meanwhile available for free as a beta product, and users can choose between three different models: Mistral Small, Mistral Large, and a prototype model called Mistral Next, designed to be brief and concise.
The company also plans to launch a paid version of Le Chat for enterprise clients, including central billing and the ability to define moderation mechanisms.
Decrypt was able to test its generation capabilities and found that it was censored, seemed competent enough, did not hallucinate excessively, had a progressive yet respectful tone, and understood long context prompts. The chatbot is not multimodal, however, and cannot access real-time information via web searches.
Founded by alumni of Google’s DeepMind and Meta, Mistral AI quickly distinguished itself in the AI sector. Within months of its incorporation in May 2023, it raised significant capital, including a $415 million funding round led by Andreessen Horowitz. Initially embracing an open-source ethos, the company has since shifted towards a business model akin to OpenAI’s, with Mistral Large being offered through a paid API.
A new partnership with Microsoft also announced today aims to extend Mistral AI’s reach by making its models available to Azure customers, a move that broadens Mistral AI’s distribution channels and could potentially help enhance its technology.
"We are thrilled to embark on this partnership with Microsoft," said Arthur Mensch, Chief Executive Officer of Mistral AI. "With Azure’s cutting-edge AI infrastructure, we are reaching a new milestone in our expansion, propelling our innovative research and practical applications to new customers everywhere."
Microsoft is OpenAI’s major shareholder, but it is also investing heavily in Open-source AI. Besides Mistral, Microsoft also has an ongoing partnership with Meta to provide the infrastructure necessary to develop LlaMA, the company’s well-known LLM.
Edited by Ryan Ozawa.