French AI startup Mistral Launching 3 new models of open weight models on Tuesday – the release of 10 models that include the largest models with multimodal and multilingual capabilities and nine models that can be used, and nine models that can be overcome.
The launch comes as Mistral, which develops open-source language models that are open and focused in Europe, has appeared to play with some of Silicon Valley’s closed-source models. The two-year-old startup, founded by a former deepmind and methods researcher, has raised $2.7 billion to date $13.7 billion worth – Frame nuts compared to number competitors Openai ($57 billion raised at a cost of $500 billion) and Anthropogenic ($45 billion multiplied by the value of $350 billion) pull.
But Mistral is trying to prove that bigger is not always better – especially for enterprise use cases.
“Our customers sometimes like to start with a large (closed) model that is not good for them … but when they install it, they are already expensive,” Guillaume Lamp and Chief, said Techcrunch. “Then they come to a nice little model to solve the use case (more efficiently).”
“In practice, the majority of enterprise use cases are things that can be applied by small models, especially if you fit in,” Lample continued.
Initial benchmark comparisons, in which smaller models trail well behind closed-door competitors, can be misleading, Lample said. Big closed source models can do better grids, but the real payoff happens when you manage it.
“In many cases, you can fit into a closed-source model,” he said.
TechCrunch events
San Francisco
I’m fat
October 13-15, 2026
The good Frontier model, called the big Mistral 3, can take some of the important capabilities of the larger AI model that corresponds to the Openai 2 and Gemini 2 opened 2, while also making a trade in Openai that is in some open competitors. Big 3 is one of the first terbetier models with multimodal and multimodal capabilities in one, it is included in llama 3 and Qwen3-Omni Alibaba 3 and QWen3-Omni. Many other companies are now pairing good large language models with smaller multi-modal models, some are doing mistrjal with models like piettral and small mistral 3.1.
The Big 3 also features an “expert mix” architecture with 41B active parameters and 675b total parameters, enabling efficient reasoning in a 256K context window. This design provides speed and capability, allowing it to process documents and function as an assistant agent for complex corporate tasks. Large Mistral position 3 suitable for document analysis, codingContent Creation, AI Assistants, and Workflow Automation.
With its new family of small models, called 3, Mistral makes the bold claim that its smaller models are not just adequate – they excel.
The lineup includes nine different models in three sizes (14B, 8B, and 3B FORICTIED), guides (models (Optimized for logical and analytical tasks).
Mistral says that this range gives developers and businesses the flexibility to model models that match the exact performance, be it raw performance, cost efficiency. The company claims that Ministrib 3 scores at par or better than other weight leaders while being more efficient and generating fewer tokens for the same task. All support variants support, handle 128k-256K context window, and can be used in languages.
A key part of the pitch is practicality. The light that works in 3, can run the GPU, which can be delivered in affordable hardware – from the game server to laptops, robots, and other peripheral devices that have limited connectivity. It is important not only for companies to stay in-house, but also for students looking for feedback or robotics teams working in remote environments. Greater efficiency, greater light, translates directly into wider access.
“This is part of our mission to ensure that AI is accessible to everyone, especially people without internet access,” he said. “We don’t want AI to be controlled by just a few big lakes.”
Several other companies are pursuing a similar efficiency trade-off model: COHO’s latest enterprise model, the A command, also only has two GPUs, and Ai Agent North Agent can run only one GPU.
Including accessibility is driving the growing focus of physical AI. Earlier this year, the company began working on integrating smaller models into robots, drones, and vehicles. Mistral collaborates with the team team agency and HTX Singapore (HTX) on special models for robotics, cyberry systems, and fire safety; With the beginning of German Defense technology Helsinki at model-action-language for drones; and with automakers Stellar in the in-car AI Assistant.
For Kawrat, reliability and independence are as critical as performance.
“Using an API from our competitor that would go down for half an hour every two weeks – if you’re a big company, you can’t afford this,” Lampu said.