French startup Mistral AI released a new large language model (LLM) called ‘Mistral Large’, which is set to rival OpenAI’s GPT-4 and Claude 2 from Anthropic, it announced on Monday.
The firm said the model “reaches top-tier reasoning capabilities” and is fluent in English, French, Spanish, German, and Italian, “with a nuanced understanding of grammar and cultural context.”
Alongside the LLM, the technology behind generative AI tools, the Paris-based company also unveiled a conversational AI chatbot called ‘Le Chat,’ or ‘the cat,’ its own alternative to ChatGPT.
Also read: Paris-based Mistral AI Secures $415 Million, Reaches $2 Billion Valuation
Mistral AI strikes deal with Microsoft
Mistral AI was founded by former Google and Meta developers in May 2023. Weeks later, the startup raised $113 million in a seed round and closed another $415 million funding round in December, which valued the company at around $2 billion.
Mistral said its large language model will be made available to Microsoft customers through the cloud computing service Azure, as the two firms revealed a partnership. Microsoft, which already hosts AI models from OpenAI on its cloud platform, will take an undisclosed stake in Mistral, according to the Financial Times.
Microsoft gained a foothold in the European AI industry with the Mistral tie-up. But the deal marks a major shift for Mistral AI, which started off with a focus on open-source models. Now, the firm has commercialized Mistral Large, just like OpenAI, and is offering the AI via a paid API, based on how frequently one uses the service.
It costs $8 per million of input tokens (basically a range of small words) and $24 per million of output tokens to ask Mistral Large questions and get a reply, TechCrunch reports.
By comparison, OpenAI’s GPT-4 Turbo costs $10 per million of input tokens and $30 per million of output tokens, the report says. This makes Mistral Large considerably cheaper than GPT-4.
“Microsoft’s trust in our model is a step forward in our journey to put frontier AI in everyone’s hands,” Mistral CEO Arthur Mensch said in a statement.
Mensch says the deal with Microsoft, which invested $13 billion in US-based OpenAI, does not mean that the company is changing its open source business model.
“We started with open source models, which anyone can deploy for free because that’s a way of distributing them widely and creating demand,” he told Le Monde.
“But from the outset, we have provided a business model with [the most powerful] optimized models.” He added that “commercial activity” will help the company fund “the costly research required for model development.”
We’re announcing a new optimised model today! Mistral Large has top-tier reasoning capacities, is multi-lingual by design, has native function calling capacities and a 32k model. The pre-trained model has 81.2% accuracy on MMLU.
Learn more on https://t.co/SOUILj8RVt.
Mistral…
— Arthur Mensch (@arthurmensch) February 26, 2024
‘Pedagogical fun’
Mistral AI said its LLM “reaches top-tier reasoning capabilities” on the most commonly used benchmarks. It said the model compares favorably with peers in reasoning and knowledge, capacity to process data or queries in many languages, as well as in completing math and coding tasks.
“It [Mistral Large] can be used for complex multilingual reasoning tasks, including text understanding, transformation, and code generation,” the company said in a blog post.
Comparison of Mistral Large with other LLMs measuring massive multitask language understanding. Source: Mistral AI.
Mistral also announced the launch of its AI chatbot, Le Chat, which the company said “is natively multilingual and offers a pedagogical and fun way to explore Mistral AI’s technology.”
Le Chat is currently available in beta to everyone for free. Users can try out three different models of the chatbot: Mistral Small, Mistral Large, and Mistral Next.
The conversational assistant is not connected to the Internet yet, Mistral said, “so it may be inaccurate or answer with outdated information.” Mistral added that it is working to make its “models as useful and as opinionated as possible.”
and then