Europe's answer to US AI dominance: Mistral's new flagship model Medium 3.5 reaches top-tier performance - with open weights and focus on data sovereignty
What it really says
French AI company Mistral AI released its new flagship model Medium 3.5 on April 29, 2026. It is a dense model with 128 billion parameters and a 256,000-token context window that combines chat, reasoning, and coding in a single model. On the SWE-Bench Verified benchmark for software engineering, it scores 77.6 percent - narrowly ahead of Anthropic's Claude Sonnet 4.5 (77.2 percent) but behind Sonnet 4.6 (79.6 percent). The model is available as open weights under a modified MIT license on Hugging Face and can be self-hosted on as few as four GPUs. API pricing is $1.50 per million input tokens and $7.50 per million output tokens - roughly half the price of comparable US models. Simultaneously, Mistral introduced Work Mode for its AI assistant Le Chat, where connections to documents, emails, and calendars are active by default. Via the Vibe CLI, developers can launch cloud-based coding agents that asynchronously handle complex programming tasks. Mistral is investing $830 million in a data center near Paris, planned to begin operations in Q2 2026.
Our assessment
This is predominantly positive news for anyone concerned about dependence on US AI providers. With Medium 3.5, a European company demonstrates it can technically compete with top models from OpenAI and Anthropic. The open-weights release is particularly significant: companies and government agencies can run the model on their own servers without sending data to external providers. This is a concrete step toward digital sovereignty - exactly what Germany's BMWE expert commission demanded just last week. The $830 million investment in European data center infrastructure underscores this commitment. Viewed soberly, there are also limitations: while Mistral is European, it is not independent of US investors - Microsoft and Nvidia are among its backers. The modified MIT license allows broad use but is not a pure open-source license. And on the latest SWE-Bench, the model still trails Sonnet 4.6. Nevertheless, competition benefits users, and a strong European provider benefits data protection.
Relevance for Germany
Directly relevant for Germany as a European alternative to US AI models. German companies hesitant to use US cloud services for AI due to GDPR concerns now have a powerful option that can run on their own infrastructure. The self-hosting capability with just four GPUs makes the model viable for mid-sized companies and public institutions. Mistral's positioning as a European alternative directly addresses the dependency concerns the BMWE expert commission articulated last week. For the German labor market, Le Chat's Work Mode is relevant: it shows how AI agents are increasingly integrated into office workflows - with all the opportunities and risks that implies for jobs and productivity. The data center near Paris could also become relevant for German companies requiring European data residency.
Fact check
Technical specifications - 128 billion parameters, 256,000-token context window, SWE-Bench Verified score of 77.6 percent - come from Mistral's official model card and Hugging Face repository. Comparison figures with Claude Sonnet 4.5 (77.2%) and Sonnet 4.6 (79.6%) are confirmed by multiple independent sources. API pricing ($1.50/$7.50 per million tokens) is documented on Mistral's platform. Self-hosting capability with four GPUs and the modified MIT license come from official documentation. The $830 million Paris data center investment was reported by Heise. Caveat: Independent third-party benchmark verification was not yet available at time of writing - scores primarily come from Mistral's own testing.
Source
- • Mistral AI Blog 29.04.2026 (mistral.ai/news/vibe-remote-agents-mistral-medium-3-5)
- • Heise 30.04.2026 (heise.de/en/news/Mistral-AI-New-Medium-3-5-Language-Model-and-Cloud-Coding-Agents-11278054.html)
- • Mistral Model Card (docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04)
- • Hugging Face (huggingface.co/mistralai/Mistral-Medium-3.5-128B)
- • The Decoder (the-decoder.com/mistrals-new-flagship-medium-3-5-folds-chat-reasoning-and-code-into-one-model/)