On April 29, Mistral AI introduced Mistral Medium 3.5, a robust model with 128 billion parameters, from its Paris lab. Despite its advanced features and agentic capabilities, the launch was met with tepid online enthusiasm.
The release included three key components: the model itself; coding agents accessible via the Mistral Vibe CLI for cloud-based, parallelized coding sessions that can automatically push pull requests to GitHub; and Work Mode in Le Chat, a consumer interface akin to ChatGPT, capable of handling complex tasks like email management, research integration, and multi-tool workflows.
Despite ambitious goals, benchmark performance reveals challenges. Medium 3.5 scores 77.6% on SWE-Bench Verified for coding accuracy and achieves 91.4% on τ³-Telecom for agentic tool use. The model consolidates weights from Medium 3.1, Magistral, and Devstral 2 into a single framework with adjustable reasoning efforts.
While unifying three models is an engineering achievement, the cost and competition are concerns. Mistral charges $1.50 per million input tokens and $7.50 per million output tokens. In contrast, Alibaba’s Qwen 3.6, which has 27 billion parameters—a fraction of Medium 3.5’s count—achieves a comparable 72.4% on SWE-Bench Verified and is available under Apache 2.0 for free use.
AI model capacity is largely determined by parameter counts; more parameters imply broader knowledge capabilities. However, open-source leaderboards show top contenders like Alibaba’s Qwen, GLM from Zhipu AI, and MiMo-V2 from Xiaomi outperforming Mistral, which has yet to rank independently.
Nonetheless, as the sole non-Chinese contender in significant open-source discussions, Mistral garners some attention. Pedro Domingos, a machine learning professor at the University of Washington, critiqued Mistral’s performance: “Only Mistral brags about how much worse its model is on benchmarks.” He questioned Europe’s role in AI advancements with a sharper inquiry: “Is it worse for Europe to not compete or to be represented by a company like Mistral?”
Youssof Altoukhi, Yoyo Studios founder, noted Qwen 3.6’s efficiency at a lower parameter count, questioning Medium 3.5’s high pricing compared to more successful competitors.
AI developer Michal Langmajer acknowledged Europe’s efforts but emphasized the need for improvement: “Their flagship model is essentially ‘not the best’ yet costs significantly more.” Some developers view open-source availability as strategic for long-term relevance rather than immediate competitive ranking, highlighting Mistral’s enterprise deployments across Europe, including a notable agreement with HSBC to self-host models.
Mistral’s appeal lies not in benchmarks but in its compliance with GDPR and suitability for European enterprises wary of non-EU data centers. AsDecrypt reported last December that HSBC chose Mistral for these reasons, valuing an EU-based provider at $14 billion despite not leading in coding or cost-efficiency. Its positioning as a non-American, non-Chinese, auditable, self-hostable option offers legal and operational security for European businesses.