News of the day
1. OVHcloud announces Olala mini, an open 7B-parameter model with no published benchmarks, launching on Hugging Face under Apache 2.0. → Read more
2. Grok Bot accessed employee accounts without permission, ignored stop orders, and denied its actions, per NewCore research. Read more here. → Read more
3. CSSF and BCL launch a 2026 questionnaire on AI use for Luxembourg financial institutions, due by 31 December 2026. → Read more
4. Nvidia chip curbs and China access test the open-source AI movement and its credibility with Jensen Huang. → Read more
Our take
Hi Dotikers!
This week, Europe got a new homegrown language model, and the story behind it is almost more interesting than the model itself. OVHcloud just announced Olala mini 7A1B, its first LLM, a compact 7 billion parameter model with only 1 billion active parameters, built on a hybrid architecture inherited from Dragon LLM, the French startup it acquired in March. It lands on Hugging Face this Thursday under an Apache 2.0 licence, trained entirely on European public supercomputers like Jupiter, MeluXina and Jean Zay.
On paper, this is exactly what Europe needs. Octave Klaba is right about one thing: training costs have dropped dramatically, and relying on the goodwill of American labs to keep releasing open models is a risky long term bet. Building that muscle at home makes strategic sense, and a model that runs at 160 tokens per second on a modest GPU, or even on a plain CPU, is genuinely useful for the kind of sovereign, low cost deployments European companies keep asking for.
But here is where it gets uncomfortable. The announcement contains plenty of speed figures and exactly zero benchmark scores. The claimed 8 GB memory footprint sits awkwardly next to weights that weigh 14 GB in BF16. And the only reference customer, Qwant, happens to belong to a company controlled by the Klaba brothers. Launching an LLM without evals is a bit like opening a restaurant and only telling people how fast the waiters walk. Fast tokens are nice, but nobody asked how smart they are.
The real test comes Thursday, when the community gets its hands on the weights. If Olala mini holds up, it will be a genuine win for European AI. If not, it will be a reminder that press releases are cheap and benchmarks are not.
Alex.
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
Meme of the day





