• Mistral AI introduces Ministral 3B and 8B for the edge.
• Benchmarks show Ministral models beat competion.

Mistral AI has introduced its latest additions to its model lineup, Ministral 3B and 8B, coinciding with the first anniversary of the well-received Mistral 7B model. The new releases are primarily intended for on-device computing and edge applications, with an emphasis on tasks like knowledge reasoning and function-calling.

The Ministral models, which can handle up to 128k context length, use a proprietary sliding-window attention mechanism that optimises inference in resource-limited environments. The models are specifically designed to fulfill the growing demand for localised, privacy-first inference in a variety of applications, including on-device translation, smart assistants, local analytics, and robotics. Acting as intermediaries for larger models, Ministral 3B and 8B improve task routing and API calling, making them adaptable in multiple use-cases.

According to Mistral AI, many of its customers and partners wanted local, privacy-driven inference solutions for critical applications such as internet-free smart assistants and autonomous robotics. “Les Ministraux were built to provide a compute-efficient, low-latency solution for these scenarios,” the company said in a blog post. “From independent hobbyists to global manufacturing teams, Les Ministraux deliver for a wide variety of use cases.”

Benchmark tests indicate that Ministral’s models outperform competitors like Gemma 2 2B, Llama 3.2 3B, and Llama 3.1 8B, making them an attractive option for enterprises focused on edge AI.

Ministral 3B and 8B are now commercially available, with prices starting at US$ 0.04 per million tokens for the 3B model and US$ 0.1 for the 8B version. The model weights for the 8B Instruct model will be offered for research purposes, making them accessible to the academic and developer communities.

Mistral AI introduced Pixtral 12B last month, a multimodal model with 12 billion parameters that can process both images and text. Pixtral, using advanced vision encoding, offers a combination of image and text processing, so it too is positioned as a prominent organisation in the emerging field of multimodal AI.

Expanding edge AI use-cases

The role of edge AI is expanding in industries like healthcare, manufacturing, and IoT, where real-time data processing and enhanced privacy are becoming important. According to Fortune Business Insights, the global edge AI market was valued at US$ 20.45 billion in 2023 and is projected to grow to US$269.82 billion by 2032,  a compound annual growth rate (CAGR) of 33.3%.

Mistral AI’s on-device models, like Ministral 3B and 8B, are designed to meet the demand for localised, low-latency AI solutions. In healthcare, for example, edge AI can be used for real-time diagnostics through wearable devices, allowing continuous patient monitoring without the need for cloud access. In manufacturing, the models can assist with predictive maintenance by monitoring machinery in real time to prevent costly breakdowns. Autonomous vehicles rely on edge AI models for the rapid decision-making required for safe operation.

By reducing the need for constant cloud communication, on-device AI models like Ministral 3B and 8B let industries process data efficiently while maintaining privacy. This capability is increasingly valuable in areas such as smart cities, robotics, and other AI-dependent applications that require real-time performance.

Sustainability and energy efficiency in AI

As AI systems advance, their environmental impact is being scrutinised. Large AI models often consume more energy due to their processing demands, especially when running in cloud environments. Mistral AI has addressed concerns with the two newly-released models, which, the company claims, are optimised for energy-efficient on-device computing. By handling complex tasks locally, the models reduce the need for energy-intensive cloud operations, which helps to achieve sustainability goals, particularly in industries where power draw is an important metric.

The energy-efficient design of Ministral models makes them suitable for sectors such as autonomous robotics and industrial IoT, where power limitations can be significant. Reducing energy consumption while maintaining high AI performance helps organisations meet both operational and environmental objectives, according to Les Ministraux.

Future outlook for Mistral AI

With US$ 640 million in recent venture capital funding, Mistral AI is positioned for further growth. The company is expected to continue growing its edge AI offerings while developing more advanced multimodal models capable of processing a wider variety of data inputs, including video and audio. Following the success of Pixtral 12B, Mistral is expected to search for new advancements in multimodal AI technology.

Mistral AI may concentrate on strengthening partnerships in industries such as healthcare, autonomous vehicles, and manufacturing, where on-device AI and privacy-first solutions are in high demand. While generating revenue remains a challenge for many AI startups, Mistral’s emphasis on edge AI may position the company to capitalise on the trends it has identified.

Want to learn more about edge computing from industry leaders? Check out Edge Computing Expo taking place in Amsterdam, California and London. 

Explore other upcoming enterprise technology events and webinars powered by TechForge here.