Mistral has expanded Mistral Studio, its enterprise AI platform. Companies can use AI agents (AI programs that carry out tasks on their own) connected to their company data and systems, along with workflows and apps, deployed in the cloud or on their own servers. Mistral also introduced its new large model, Mistral Large 4.
Agents can be built to investigate and carry out multi-step tasks on their own.
Developers can deploy what they build with technical tools to the Studio environment with a single command.
Companies can control tools, data, and actions by setting their scope.
Some customers, including Spain's ABANCA and France's BNP Paribas, are already using Studio in real business work.
French AI company Mistral has released the weights (the learned numerical values that make up a model) of its large AI model Mistral Large 4, nicknamed Le Chonk. It was first previewed last month, and reinforcement learning (training that builds skill through trial and error) was added to improve its performance. The company says it is the best publicly available model for coding.
It earned the highest score among open models in coding evaluations, and its score on terminal task evaluations rose 13 percentage points to 41%.
It greatly reduces the number of tokens (units of text that AI reads and writes) needed to solve the same tasks, lowering running costs.
It offers three reasoning modes, low, medium, and high, so users can choose how deeply the model answers.
Mistral has released Voxtral 3, a new set of models for building AI voice agents that can listen, think, and speak. It includes a speech-to-text model, a text-to-speech model, and a software development kit (SDK) that links them together. It can be used through servers in Europe, and companies can also install it directly on their own servers.
The speech transcription model handles more than 50 languages, can tell who is speaking, and can help it better recognize specific words.
The speech synthesis model speaks naturally by changing emotion, speaking speed, and emphasis to suit the conversation.
Voice agents can look up information during a conversation and operate company systems to complete tasks over the phone or on the web.
The Pro Voice feature for experts can reproduce one person's speaking style and emotional expression in multiple languages using only 30 minutes of audio.
Mistral has released Timestral, a set of AI models that predict time-series data. The largest model has 209 million parameters (the internal numbers an AI adjusts during training), and it ranked first on three public benchmarks. The models come in four sizes, from 5 million to 209 million parameters.
Time-series data is information collected in time order, such as sensor readings or server metrics.
The models are pre-trained, so they can predict on data they have never seen before without extra training.
The smaller models reportedly run quickly on ordinary server CPUs without a GPU.
The company says that at each size, the models are the most accurate, even compared with larger models.
French AI company Mistral has started a public preview of its new large AI model, ML4. Developers and companies can try it now through the preview API, and the model weights (the numerical results of training the model) will be released later this month. ML4 is a multimodal model that understands both text and images.
ML4 has 100 billion total parameters, but only 52 billion are used in each actual calculation.
It was trained from scratch on 3,800 Nvidia GPUs in Mistral's own European data centers.
The company says it is the best among public models in security, finance, and law.
Security experts and partners are first testing a version with fewer restrictions.
Mistral and Mozilla are partnering to put Mistral's AI models into Smart Window (beta), Firefox's AI browsing assistant. It will first be offered to users in France and North America, with plans to expand to the UK and Germany this year.
Smart Window organizes complex search results, remembers pages you closed by mistake, and finds needed information based on the tabs you have open.
By default, conversations are not stored on Mozilla's servers, and Mistral has promised not to store data either.
The AI is being refined to reflect regional languages, dialects, and cultural context, so people can get answers that understand the tone and nuances of their own region.