EU-AI Monthly Update #2: System One Models, Mistral is back, and European Jev is here.

11 October 2026 - Written by M.J (not AI assisted):


Welcome to another monthly report from EUAINews, your daily source for EU-AI News via LinkedIn or Reddit and your monthly source to quickly catch up on all things EU-AI.

Let's start with the biggest foundational EU-AI movers:

Groundbreaking EU-AI changes (2)

1. Mistral Large 4

Mistral launched Mistral Large 4, "Le Chonk", achieving a cost-efficient and competitive European model (currently in public preview). Benchmarks show significant progress compared to past Mistral models.

Describe Image Here

Mistral Large 4 VS Mistral 3.5 Medium + SOTA Models (higher is better) - source


On the global stage, however, Mistral is still behind the latest US and Chinese models:

Describe Image Here

However, the bigger picture is perhaps more interesting: Mistral has shown significant improvement in benchmarks compared to their own past top model, indicating that they may keep improving at a similarly steep rate in the future.

Describe Image Here


2. System One Models & EU Jev Alternative: Strom

Europe gets a Jev alternative: 🇩🇪 Uprelic launches Strom for system one decision-making, including decisions based on images, cutting costs by 10-100X for simpler workflows. Their move was part of a larger trend: System one decision models.

What are system one models?

System one decision models like TypeSafe's Jev (currently valued at $7.5B) skip many layers that traditional LLMs use to continuously determine the next token.

Instead of generating text, system one models convert input to a simple number which can be used to pick one option (e.g. choice 1,2,3 or 4), a yes/no answer (noul) or a score between 0 and 1.

Describe Image Here

As a result of this simplicity, input tokens cost about 10-100x less than those of traditional LLMs, output is free, and decisions can be made in milliseconds instead of seconds, unlocking interesting new use-cases, such as classifying large datasets or working with realtime data.

Besides text-based input, Uprelic also showed that Strom can read images natively, beating TypeSafe's Jev in a simple flight booking example, where a web browser was used to automatically book a flight.

Describe Image Here

Significantly Improved AI Capabilities (2)

  • 🇩🇪 Visuals: Black Forest Labs releases Flux 3 Image, bridging the gap between 'low-quality AI slop' and 'precise high-quality image generation' in 4K.

Describe Image Here

  • Voice: 🇬🇧 ElevenLabs introduces "the next frontier of human-level communication", hyper-realistic (emotional) voiceovers.

Other significant stories: