Just three and a half months after securing substantial funding as a fledgling startup, French company Mistral AI has introduced its first major open-source language model, Mistral 7B. With this release, the company aims to compete with top-performing generative AI language models (LLMs).
Mistral AI Unveils Mistral 7B: An Open-Source Language Model
On September 27, Mistral AI made headlines with an unexpected tweet sent out from its official account at 5 a.m. The tweet included a magnet link, accompanied by alphanumeric characters resembling a URL. Among them, the standout was “mistral-7B-v0.1,” signifying the name of the French startup’s first significant language model. This announcement was later made more traditionally in a blog post.
Mistral 7B is, as the name suggests, a 7.3-billion-parameter LLM. It’s open source, aligning with the promise made by Mistral AI’s founders, who raised €105 million just a few months ago. This funding came hot on the heels of their company’s notable launch in the midst of generative AI frenzy. Described as the outcome of three months of intense work, the model unveiled is positioned by its creators as outperforming all open-source competitors in its category.
More Powerful Than Llama 2
Mistral AI places significant emphasis on its rivalry with Meta. Mark Zuckerberg’s company introduced its open-source Llama 2 model in July, and it now faces robust competition in the open generative AI arena. According to calculations by the French startup, its Mistral 7B model surpasses not only Llama 2’s 7B and 13B versions but also outperforms Llama 1’s 34B version across many benchmarks.
Arthur Mensch, CEO of Mistral AI, previously labeled the release of Llama 2 as an “interesting challenge.” Now, his company can confidently claim its place among the frontrunners. The startup’s leadership hopes that the adoption of their model will match their ambitions. Published under the open-source Apache 2.0 license, Mistral 7B is accessible via platforms like GitHub and Hugging Face, enabling anyone to use it freely. To facilitate communication, collaboration, and support questions regarding this innovation, a Discord server has been opened.
Open Source as a Guiding Principle
Mistral AI already announces the training of “much larger models” and mentions “new architectures.” As the startup aims to gradually close the gap with the leaders in generative AI, such as OpenAI, it remains firmly committed to open source. In a blog post from September 27, the company stated, “We believe that by training our own models, openly publishing them, and encouraging community contributions, we can build a credible alternative to the emerging AI oligopoly.”
Mistral AI’s leaders bank on the “demand for technical excellence” inherent in the open-source world, as well as the flexibility and transparency enabled by open access to the code. Developers have complete control over the engine powering their applications, enabling them to adjust the model’s size to fit their needs while managing costs and latency. Additionally, open models serve as valuable safeguards against the misuse of generative AI, allowing public institutions and private companies to identify vulnerabilities and detect improper usage.
Commercial Offerings in the Pipeline
To drive practical use of its cutting-edge model and establish a foothold among businesses in the burgeoning generative AI market, Mistral AI intends to develop a commercial offering in the future. The blog post states, “We will offer proprietary models optimized for deployment on-site and in private clouds. These models will be distributed as ‘white box’ solutions, providing both weights and source code.”
In this context, Mistral 7B will directly compete with Meta’s open-source LLMs and OpenAI’s paid ChatGPT Enterprise, which provides access to the latest large language models. Beyond Western actors, Mistral AI will also keep a watchful eye on the Chinese AI ecosystem, where Alibaba Cloud launched several open-source models in both Chinese and English in August, also boasting 7 billion parameters.