Recently, French AI startup Mistral AI announced their new generation of language models — Ministral3B and Ministral8B.
These new models are part of the "Ministraux" series, designed specifically for edge devices and edge computing scenarios, supporting a context length of up to 128,000 tokens. This means that these models not only have powerful processing capabilities but can also be used in situations where data privacy and local processing are particularly important.
Mistral states that the Ministraux series models are well-suited for a range of applications, such as local translation, offline smart assistants, data analysis, and autonomous robotics. To further enhance efficiency, the Ministraux models can also be combined with larger language models (like Mistral Large) as effective intermediaries in multi-step workflows.
In terms of performance, Mistral's benchmark tests show that Ministral3B and 8B outperform many同类models in multiple categories, such as Google's Gemma22B and Meta's Llama3.18B. Notably, despite having fewer parameters, Ministral3B outperforms its predecessor Mistral7B in certain tests.
In fact, Ministral8B excels in all tests, particularly in knowledge, common sense, function calling, and multilingual capabilities.
Regarding pricing, these new models from Ministral AI are already available via API. Ministral8B costs $0.10 per million tokens, while Ministral3B is $0.04. Additionally, Mistral offers the model weights for Ministral8B Instruct for research purposes. It is worth noting that these new models from Mistral will soon be available through cloud partners like Google Vertex and AWS.
Key Points:
- 🚀 Mistral AI introduces Ministral3B and 8B, supporting up to 128,000 tokens in context length.
- 💡 These models are suitable for local translation, offline assistants, data analysis, and autonomous robotics.
- 💰 Pricing: Ministral8B costs $0.10 per million tokens, Ministral3B is $0.04.