A small model, described in unusual detail
OVHcloud's release specifies the parameter count, the active parameter count, the number of experts per layer, the two attention mechanisms being combined, and the numeric format. It names the three public supercomputers the model was trained on and puts a value on the public support behind them: €7 million in computing hours.
That level of detail makes the omission easier to spot. Nowhere in the release is there a score on any standard benchmark. There is no reasoning result, no coding result, no language test, nothing a reader could set beside the small open models already available at this size. A technical report is listed in the planned release package, and it may contain those numbers when it can be read. The announcement itself does not.
Speed is the only performance on offer
What the release offers instead of quality scores is speed: 160 tokens per second on an RTX A4500, 50 tokens per second on a standard processor, a fit inside 8 GB of GPU memory. Those are useful claims for a model aimed at workstations and local deployment, and the 1-billion-active design is a plausible route to them. But throughput depends on measurement conditions, which this release does not supply: it does not say what prompt length, output length or batch size produced the figures.
The one production result has the same shape. OVHcloud says Qwant, the European search engine, runs Olala mini for its flash answers, with answers 20% faster and running costs divided by three. Divided by three compared with what is the obvious question, and the release does not answer it: the model Qwant used before is not named. Without the predecessor and the workload, readers cannot tell how large the efficiency gain is, and neither can prospective users.
The strategic point is real
None of this makes the model unimportant. OVHcloud's chief executive, Octave Klaba, states the motive plainly in the release: the company does not want to depend on other labs continuing to share open models, and training costs have fallen sharply (Klaba puts the fall at a factor of 20 in three years). ALFW reads the announcement as an effort to reduce dependence on external model suppliers: OVHcloud says it serves 1.7 million customers and has internal uses, from its own workspace products to DevOps and security tooling, where a small model of its own is the point. The announced Apache 2.0 license, the AI Act documentation listed in the release package, and training entirely on European public machines are all consistent with that pitch. Deployment economics remain unquantified.
The roadmap also tells readers how to weight the mini. Version 1.1, expected by the end of the year, would stretch the context window from 32,000 to 128,000 tokens and improve mathematical reasoning. Version 2.0 is a company target above 100 billion parameters. Olala mini is a proof that OVHcloud can train a model, sized for adoption, not a bid for the top of a leaderboard. It would be easier to say that with confidence if the company had published the scores that would show it.
Sources
- [1] OVHcloud, "OVH announces Olala mini 7A1B", press release, November 17, 2026 (read first-hand in full by ALFW; full-text check: no benchmark scores, no test conditions)Read source
- [2] Unite.AI, "OVHcloud Introduces Olala Mini 7A1B as Its First Open-Source LLM", November 17, 2026 (secondary coverage)Read source