senn-techAI Radar: Mistral Small 4 & Voxtral — Europe's Open Answer
Out of Europe comes Mistral, the lab to watch when data sovereignty means not just technology but also jurisdiction. Its recent releases pack a lot into a little — and at the same time show where you need to read the license closely.
What's New
- Mistral Small 4: 119B total parameters, around 6B active, 256K context — and for the first time unifying reasoning, multimodal processing, and agentic coding in one model. Under Apache 2.0.
- Mistral 3 / Large 3: the dense 3B, 8B, and 14B models plus Large 3 (41B active / 675B total) — the largest open MoE from a major lab, also Apache 2.0.
- Voxtral TTS: Mistral's first speech model (text-to-speech), nine languages including German — a direct shot at ElevenLabs.
The Caveat
Important for business: Voxtral TTS is licensed CC BY-NC 4.0 — non-commercial use only. You may self-host it for research and testing, but production use requires a license through Mistral. Only the speech-recognition model, Voxtral Realtime, is Apache 2.0 and thus commercially free. The text models (Small 4, Large 3), by contrast, are clearly Apache-licensed.
Our Take
Mistral remains, for us, the obvious first stop when data sovereignty and EU jurisdiction line up. Small 4 is an honest workhorse for self-hosting; with Voxtral, it pays to read the fine print before a voice ends up in a product. European does not automatically mean carefree — but it is a good place to start.