Mistral AI Drops New Open-Source Model. The Internet Is Not Impressed, Except for One Thing

Mistral AI released Mistral Medium 3.5 on April 29. The Paris-based lab announced a compact model with 128 billion parameters, a set of agentic functions – and ran straight into a wall of online “meh” responses.

The release consisted of three parts. Firstly, the model itself. Second, remote coding agents via Mistral Vibe CLI: cloud-based coding sessions that can push pull requests to GitHub and run in parallel without having to sit at a terminal. Third, Work Mode in Le Chat, Mistral’s ChatGPT-style consumer interface, which now handles multi-step autonomous tasks such as email triage, research synthesis, and cross-tool workflows.

Big ambitions, but a messy benchmark reality.

Medium 3.5 scores 77.6% on SWE-Bench Verified – a coding benchmark that tests whether a model can solve real GitHub problems by generating working patches. It also scores 91.4% on τ³-Telecom, which measures the use of agentic tools in specialized environments. Mistral has also merged three previously separate models (Medium 3.1, Magistral and Devstral 2) into a single set of weights with configurable reasoning efforts per request.

A unified model that replaces three is a real technical victory. The problem is what it costs and who it affects.

Mistral charges $1.50 per million input tokens and $7.50 per million output tokens. Alibaba’s Qwen 3.6, with 27 billion parameters (less than a quarter of the number of parameters of Medium 3.5), scores 72.4% on the same SWE-Bench Verified benchmark and ships under Apache 2.0, meaning you can download and run it for free.

Scroll through the open source rankings and the picture is grim. The top spots belong to Alibaba’s Qwen, GLM from China’s Zhipu AI and MiMo-V2 from Xiaomi, all cheaper, more powerful and more competitive than Mistral’s new release. Medium 3.5 isn’t even on the major independent rankings yet; third-party evaluations are still pending.

The one good thing, however, as some claim, is that Mistral is the only non-Chinese model with any serious presence in the open source conversation at the moment.

The Internet responds

Pedro Domingos, professor of machine learning at the University of Washington, was not gentle:

“Main AI companies brag about how much better their model is on benchmarks. Only Mistral brags about how much worse its model is.”

He continued with a more pointed question: “I don’t know what’s worse: Europe not joining the AI ​​race or being represented by a laughing stock like Mistral.”

Youssof Altoukhi, founder of Yoyo Studios, did the math: With 27 billion parameters, Qwen 3.6 is 4.7 times smaller than Medium 3.5 and scores comparable in terms of coding. Medium 3.5’s output prices place it alongside closed models that score significantly higher on every major benchmark.

“Without their political skill, they would have been bankrupt by now,” he said.

Not everyone was purely dismissive. AI developer Michal Langmajer summarized the ambivalence:

“I’m really happy that there is still a non-American, non-Chinese lab trying to build groundbreaking LLMs, but we need to step up the game in Europe. Their new flagship model is basically ‘not the best’ on any benchmark, but costs several times more than most competitors.”

Some developers argued that open weights are a sustainability game and not a ranked game. Today, a model that anyone can download, refine, and self-host doesn’t need to win rankings to stay relevant. Others pointed to real enterprise deployments of Mistral across Europe as evidence that the moat is not purely technical.

The geopolitical safety net

This is where the actual field of Mistral lives.

European companies subject to GDPR, banks that process sensitive customer data, and governments that don’t want to route AI workloads through Chinese infrastructure have limited options. If Declutter Last December, HSBC announced a multi-year agreement with Mistral, specifically for self-hosting models on its own infrastructure. The appeal of an EU-based open-weight lab with a $14 billion valuation doesn’t show up in benchmark tables, but it does in procurement decisions.

Not the best at coding, and not the cheapest. But it is: not American, not Chinese, verifiable, self-hostable and legally safe for European companies.