The large language model Mistral 7B is free to download, with no restrictions on use.
French generative AI startup Mistral AI has released a free-to-use large language model called Mistral 7B, which has 7.3 billion parameters and has proven competitive with Meta’s Llama 2.
Mistral 7B can be downloaded directly from Mistral AI’s blog post.
Jump to:
Mistral AI positions its LLM as different from the competition like LLama 2 and GPT-4 because it is open and community-focused instead of proprietary. (Meta calls Llama 2 open source.) Mistral 7B can be downloaded for free, and Mistral AI set up a Discord channel and a GitHub page for community discussion. Mistral 7B is released under the Apache 2.0 commercial use license, meaning there are no restrictions on how it is used or by whom.
“Community-backed model development is the surest path to fight censorship and bias in a technology shaping our future,” stated the Mistral AI team in a blog post on September 27.
Being able to adjust model sizes and costs will make generative AI more affordable and reduce latency, Mistral AI said.
Mistral AI claims Mistral 7B “outperforms all currently available open models up to 13B parameters on all standard English and code benchmarks,” including Llama 2. Grouped-query attention speeds up inference, while sliding window attention handles long sequences cost effectively, Mistral AI said.
“This is the result of three months of intense work, in which we assembled the Mistral AI team, rebuilt a top-performance MLops stack and designed a most sophisticated data processing pipeline, from scratch,” the Mistral AI team said in the blog post.
While Mistral AI hasn’t announced statistics relative to OpenAI’s GPT-4, it does seem to be throwing its hat into the same ring.
Like other generative AI, Mistral 7B can generate text (such as emails, presentations or project outlines), write code, summarize text and answer questions. Mistral AI proposes the Mistral 7B LLM is a way to combat AI-based misinformation. “They [open generative AI models] will allow public institutions and private companies to audit generative systems for flaws, and to detect bad usage of generative models,” Mistral AI wrote in a blog post.
Mistral AI anticipates businesses across industries will want to create generative AI models that specialize in the business’ own tasks. Compressing each model and connecting it to specific modalities will help tailor Mistral 7B for efficient use, Mistral AI said.
SEE: Expectations are inflated around generative AI as companies search for ways to make generative AI products productive. (TechRepublic)
Mistral AI plans to offer a paid version of its generative AI in the future; the paid version will offer a “white box” solution with training weights and code sources open to the user. The company anticipates offering hosted solutions and dedicated enterprise deployment at an unspecified later date.
Mistral AI cofounder and CEO Arthur Mensch previously worked as a research scientist at Alphabet DeepMind; Chief Scientist and cofounder Guillaume Lample was a research scientist at Meta (known at the time as Facebook) AI research; and CTO Timothee Lacroix worked as a research engineer at Meta.
Mistral AI received $113 million in seed funding from Lightspeed Venture Group in June.
