Mistral, an AI startup based in Paris, has released two new AI models specifically optimized for running on local devices. These models, called Ministral 3B and Ministral 8B, are designed to work efficiently on devices like laptops and smartphones without needing a constant connection to cloud servers. The idea behind this move is to meet the rising demand for privacy-focused and low-latency AI solutions.
Targeting Local AI Processing
The demand for AI models capable of running on personal devices has grown, as companies look for ways to ensure more secure, on-site data processing. Mistral says the new models can be applied in areas such as translation services that don’t need internet access, local analytics, and robotics. By processing data locally, these models enable users to keep everything in-house, minimizing the need for cloud-based processing.
Mistral says that their customers have been pushing for options that don’t rely on cloud infrastructure but still offer rapid response times. The new models provide just that, running efficiently on edge devices.
Specifications and Pricing
Ministral’s newest models both offer a 128,000-token context window, which is about the same as being able to process a 50-page document all at once. For developers and businesses, the larger of the two, Ministral 8B, is available for research today, with pricing of 10 cents per million tokens. The smaller 3B version costs 4 cents for the same number of tokens, making it accessible for smaller-scale operations or developers just starting out with AI integration.
Both models will soon be available for deployment through Mistral’s cloud service, “Le Platforme,” as well as other popular cloud partners.
Ministral 3B and 8B base models compared to Gemma 2 2B, Llama 3.2 3B, Llama 3.1 8B and Mistral 7B
Small, Efficient Models Gain Popularity
As AI models continue to evolve, there’s been an increasing shift towards smaller, more efficient options. These types of models tend to be less expensive to train and faster to run, compared to larger AI models. Mistral is not alone in this space; competitors like Google have developed their own “Gemma” models, while Microsoft has its own collection of small models known as “Phi”.
Mistral claims that their models outperform similar offerings from both Google and Microsoft, particularly in tasks like following instructions and problem-solving. They also outperform Mistral’s own older models, showing the company’s focus on improving efficiency without compromising on capability.
A comparison of the 8B family of Instruct models – Gemma 2 9B, Llama 3.1 8B, Mistral 7B and Ministral 8B
Beyond Just Text Generation
Though primarily designed for text-based applications, Mistral’s models have been adapted for broader uses. Whether it’s parsing input data or managing complex workflows, the models can serve as intermediaries between larger AI systems. This flexibility allows them to handle everything from user commands to API calls without slowing down operations. In sectors like robotics, these on-device AI systems could enable smarter, more autonomous machines.
Growing Product Line and Investment
Mistral’s rapid growth is being fueled by significant investment. After raising $640 million in venture capital, the company has been steadily expanding its product offerings. Along with its latest models, Mistral has rolled out a software development kit (SDK) for customers who want to fine-tune the AI models for their own unique needs.
The company’s product portfolio already includes tools for developers to test AI systems and a new model specialized in writing code, called Codestral. As the company’s technology continues to develop, it’s clear that Mistral is aiming to offer versatile, local AI options for a range of industries.
Privacy-First AI for All Devices
Ministral’s focus on privacy and local computing isn’t just about keeping data secure—it’s also about offering fast, responsive AI that doesn’t rely on an internet connection. For industries dealing with sensitive information, this could be a game-changer. From translation services that work offline to analytics tools that don’t send data to the cloud, these AI models have the potential to make AI more accessible and secure for a wide range of use cases.
The availability of these new models could encourage even more companies to look into on-device AI, potentially shifting the market away from heavy cloud dependency.