Mistral, a French AI startup, has introduced its first generative AI models designed for use on edge devices like laptops and smartphones. The new models, known as "Les Ministraux," are intended for a range of tasks, from basic text generation to more advanced operations when paired with other models.
There are two versions of these models: Ministral 3B and Ministral 8B. Both come with a context window of 128,000 tokens, meaning they can process about the length of a 50-page book at once. These models are especially useful for applications where local, private processing is key, such as on-device translation, offline smart assistants, and local data analysis.
Mistral explains that the demand for privacy-first AI solutions on devices has grown, particularly in areas like autonomous robotics and real-time analytics. The company designed Les Ministraux models to meet this demand with efficient computation and low latency.
For now, Ministral 8B is available for download, but only for research purposes. Developers and companies wanting to use it for commercial purposes must contact Mistral for a license. The models can also be accessed via Mistral’s cloud platform, "La Platforme," or other cloud services in the near future. Ministral 8B is priced at 10 cents per million tokens, while Ministral 3B costs 4 cents per million tokens.
This development reflects a broader trend in AI toward smaller, more efficient models. These are quicker to train, cheaper to deploy, and suitable for edge devices. Companies like Google and Microsoft are also releasing small AI models, with Google’s Gemma and Microsoft’s Phi families leading the way. Meta has also released smaller models optimized for edge hardware with its Llama suite.
Mistral claims that its Ministral 3B and Ministral 8B outperform comparable models like Meta’s Llama and Google’s Gemma on tasks related to instruction-following and problem-solving.
Having raised $640 million in venture capital, Mistral continues to expand its AI offerings. It recently introduced a free service for developers to test its models, an SDK for fine-tuning, and a generative code model called Codestral. Founded by former employees of Meta and DeepMind, Mistral aims to develop flagship AI models that can compete with top-tier models such as OpenAI's GPT-4 and Anthropic's Claude, while also striving to build a sustainable business model.
