French AI startup Mistral has released the first generative AI model designed to run on edge devices such as laptops and mobile phones.
This new family of models, which Mistral calls “Les Ministraux,” can be used or adapted for a variety of applications, from basic text generation to working with more capable models to complete tasks. You can.
Two Les Ministraux models are available: Ministral 3B and Ministral 8B. Both have a 128,000-token context window and can capture approximately the length of a 50-page book.
“Our most innovative customers and partners are increasingly demanding local, privacy-first inference for critical applications such as on-device translation, internet-free smart assistants, local analytics, and autonomous robotics. ,” Mistral wrote in a blog post. “Les Ministraux was built to provide a computationally efficient, low-latency solution for these scenarios.”
Ministral 8B is available for download as of today, although strictly for research purposes. Mistral is asking developers and companies interested in self-deployment setups for Ministral 8B or Ministral 3B to contact Mistral to obtain a commercial license.
Otherwise, developers will be able to use Ministral 3B and Ministral 8B through Mistral's cloud platform, Le Platforme, as well as other clouds the startup partners with in the coming weeks. Ministral 8B costs 10 cents per million output/input tokens (approximately 750,000 words), while Ministeral 3B costs 4 cents per million output/input tokens.
Recently, there has been a trend toward smaller models that are cheaper and faster to train, fine-tune, and run than larger models. While Google continues to add models to the Gemma small model family, Microsoft offers the Phi collection of models. In the latest update to the Llama suite, Meta introduced several smaller models optimized for edge hardware.
Mistral found that the Ministral 3B and Ministral 8B outperformed comparable Llama and Gemma models, as well as the proprietary Mistral 7B, on several AI benchmarks designed to assess command-following and problem-solving abilities. I claim that there is.
Paris-based Mistral recently raised $640 million in venture capital and continues to gradually expand its AI product portfolio. Over the past few months, the company has launched new models, including a free service for developers to test models, an SDK for customers to fine-tune those models, and a code generation model called Codestral.
Co-founded by Meta and Google's DeepMind alumni, Mistral's stated mission is to create a flagship model that rivals today's best-performing models, such as OpenAI's GPT-4o and Anthropic's Claude, and ideally is to make a profit in the process. While the “making money” part has proven to be difficult (as with most generative AI startups), Mistral reportedly started turning a profit this summer.