Mistral announced a strategic collaboration with HUMAIN (Saudi Arabia) worth hundreds of millions of Euros, covering AI infrastructure, model development, and deployment across the Middle East. The partnership includes development of Arabic-language frontier models and localized models for cybersecurity and voice use cases. Mistral will explore using HUMAIN's data center infrastructure to support regional compute needs, with a joint go-to-market strategy targeting regulated industries.

Mistral introduced Agentic Search, a multi-step retrieval loop system that allows models to search, navigate, and verify information across complex data sources. It replaces traditional one-shot RAG with iterative retrieval using five tools: search, open, navigate, read, and grep. It delivers up to 3x correctness on financial filings and reduces p90 latency by up to 39.6%.

Mistral's platform now supports third-party open models, launching with Z.ai's GLM-5.2, running under the same regional infrastructure, data-residency controls, and SLA commitments as native Mistral models. Simultaneously, Mistral launched generally available Regional Endpoints (EU/US choice) and a new Priority Tier in public preview with committed rate limits and uptime SLAs for mission-critical workloads.

Mistral released Shieldstral, a 3B open-weights multimodal safety classifier that frames content moderation as a policy-adaptive question-answering task. It accepts plain-language policies at inference time without retraining, unifying text and image safety evaluation in a single model. Released under Apache 2.0, it outperforms models up to 7x its size on text safety benchmarks and sets a new state of the art on multimodal moderation.

Mistral AI introduced Robostral Navigate, an 8B model built for embodied robotic navigation that uses only a single RGB camera and no depth sensors. It achieves 76.6% success rate on the R2R-CE validation unseen benchmark, outperforming multi-sensor approaches. The model uses a pointing-based navigation technique combined with reinforcement learning and was trained entirely in simulation on 400,000 trajectories across 6,000 scenes.

Leanstral 1.5 is a free Apache-2.0 licensed model with 119B total and only 6B active parameters focused on formal verification in Lean 4. It saturates miniF2F completely (100%), solves 587/672 PutnamBench problems, and achieves new state-of-the-art results of 87% on FATE-H and 34% on FATE-X. It also uncovers previously unknown bugs in open-source repositories and demonstrates strong test-time scaling behavior.

Mistral OCR 4 introduces bounding boxes, typed-block classification (titles, tables, equations, signatures), and inline confidence scores alongside extracted text. The model supports 170 languages across 10 language groups and can run in a single container for fully self-hosted deployments. It achieves top performance on OlmOCRBench (85.20) with human evaluators preferring it over competing systems at an average 72% win rate.

Mistral announced 'Vibe', a unified agent for long-horizon productivity that handles inbox/calendar management, deep research, drafting deliverables, and orchestrates recurring workflows. It also handles coding tasks end-to-end, from building features and fixing bugs to shipping reviewable pull requests. Vibe runs on flagship Mistral models optimized for reasoning, agentic tasks, tool calls, and coding.

Le Chat has been rebranded and evolved into Vibe, a unified agent for both long-running work tasks and coding. Vibe introduces Work Mode for complex multi-step tasks across enterprise tools like Google Workspace, Slack, and GitHub, and Code Mode for remote coding agents that can build features, fix bugs, and ship pull requests. A new Mistral Vibe extension for VS Code was also released, bringing the coding agent directly into the IDE.

Mistral AI introduced Mistral Small 4, a new model version. Additionally, Voxtral TTS was released as a frontier, open-weights text-to-speech model that is fast, instantly adaptable, and produces lifelike speech for voice agents. Mistral also introduced Forge, a system for enterprises to build frontier-grade AI models grounded in proprietary knowledge.

Mistral introduced Mistral 3, described as a family of frontier open-source multimodal models. This represents a new generation of models with multimodal capabilities. It is positioned as a significant step forward in both openness and frontier-level performance.

Significantly improved coding and reasoning. 128K context window. Better instruction following and reduced hallucinations.

Flagship commercial model. Top tier reasoning capabilities. Native function calling. Available via API and Azure.

Mixture of Experts architecture. Matches GPT-3.5 performance at fraction of the cost. Multilingual support across 5 languages.

First Mistral model released. Outperformed Llama 2 13B on all benchmarks despite being smaller. Apache 2.0 license for commercial use.