Mistral AI has launched Mistral 3, a collection of 10 open-source AI models designed to run on devices ranging from smartphones to enterprise cloud systems. The French startup released all models under the Apache 2.0 license, allowing unrestricted commercial use.
The release includes Mistral Large 3, the company’s flagship model, and the Ministral 3 series comprising nine smaller models optimized for edge computing. Mistral Large 3 uses a mixture of experts architecture with 41 billion active parameters drawn from a total pool of 675 billion parameters. The model processes both text and images and handles context windows up to 256,000 tokens.
Mistral trained Large 3 on 3,000 NVIDIA H200 GPUs. According to the company, the model ranks second among open-source non-reasoning models on the LMArena leaderboard. Mistral emphasizes that Large 3 demonstrates particular strength in multilingual conversations beyond English and Chinese.
The Ministral 3 lineup offers three model sizes: 3 billion, 8 billion, and 14 billion parameters. Each size comes in three variants serving different purposes: base models for extensive customization, instruction-tuned models for general tasks, and reasoning-optimized models for complex logic. All variants include image understanding capabilities.
Guillaume Lample, Mistral’s chief scientist and co-founder, told VentureBeat that the smallest Ministral 3 models can run on devices with as little as 4 gigabytes of video memory using 4-bit quantization. This enables deployment on standard laptops, smartphones, and embedded systems without cloud infrastructure or internet connectivity.
Mistral’s approach differs from competitors like OpenAI, Google, and Anthropic, which focus on building larger proprietary systems. Lample explained that the company targets enterprise customers seeking alternatives to expensive closed-source models. He said that in more than 90 percent of cases, a fine-tuned small model can outperform larger general-purpose systems on specific tasks.
The company positions the release as addressing practical business concerns. Lample noted that customers often build prototypes with closed-source models but find deployment costs prohibitive at scale. Mistral offers engineering teams that work directly with customers to create synthetic training data and fine-tune smaller models for specific workflows.
Mistral 3 is available on multiple platforms including Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, and several other services. The company also provides custom model training services for organizations requiring tailored solutions.
The release occurs amid competition from Chinese companies like DeepSeek and Alibaba’s Qwen series, which have advanced their open-source offerings. Mistral differentiates itself through multilingual capabilities and integrated multimodal processing within single models rather than separate systems.
Mistral has raised approximately 1.05 billion dollars in funding since its founding in May 2023. The company was valued at 6 billion dollars in June 2024 and more than doubled its valuation in a September funding round that included investment from ASML, the Dutch semiconductor equipment manufacturer.
The company has expanded beyond model development to offer a comprehensive enterprise platform including AI Studio for production deployment, the Mistral Agents API, and Le Chat assistant with research and voice capabilities.
Sources: Mistral, VentureBeat
