AI Momentum 107.6 +2.7%GitHub Pulse 5172 -5.9%Capital Flow $204.6B +5804.2%Hiring Heatmap 0.0 -100.0%
Advertisement
Open Source

French cloud company OVHcloud gives away its first AI model, saying it can't rely on other labs to keep sharing theirs

OVHcloud, a French cloud provider with 46 data centers and about 1.7 million customers, published its first AI model on Thursday. It is free to download, change and use commercially under Apache 2.0, a standard open-source license with no revenue caps or regional bans.

The model, Olala mini, was unveiled on Tuesday and formally launched at the company's AI Summit in Paris. Chief executive Octave Klaba said "we cannot simply rely on the generosity of certain labs" to keep releasing their models openly in the years ahead.

For buyers, a European supplier with a plain license reduces one source of risk. Whether the model is good enough to matter is a separate question.

What OVHcloud is offering

Olala mini is small by industry standards. It holds about 7 billion parameters, the internal settings that store what a model has learned. It is built as a mixture of experts: only about 1 billion of those settings are switched on for each request. That makes it cheaper and faster to run.

OVHcloud says it reaches 160 tokens (word fragments) per second on a mid-range Nvidia workstation card and 50 per second on an ordinary processor with no graphics chip. It can read about 32,000 tokens at once, roughly a long report. A version handling four times that is promised by the end of the year.

It was trained entirely on three public supercomputers in Germany, Luxembourg and France. OVHcloud values the public support at about €7 million in computing hours. It has not said what the model cost in total, or how those hours were allocated.

The release also includes paperwork for the EU AI Act, Europe's new AI law, including a public summary of the training material. That is a selling point for banks, hospitals and governments that must document the AI they use. OVHcloud has not published a price for running the model on its own cloud service.

The roadmap is bigger than the product. OVHcloud says a version 2.0 will exceed 100 billion parameters, with specialized models for business automation, cloud operations and cybersecurity. It gave no date.

The evidence so far

When the model was announced on Tuesday, OVHcloud published speed figures but no quality scores. Trade outlets including ActuIA and byteiota flagged that gap. The model's download page, posted Thursday, now carries a scorecard. OVHcloud ran every test on it itself, and no outside evaluator has published scores yet.

On that scorecard, Olala mini averages 68.0 across twelve tests of instruction following, reasoning, coding and math. OVHcloud says it is strongest at following instructions and weakest at competition math. OVHcloud does not average the rival models, but by our calculation from its table, Alibaba's Qwen 3.5 4B, a model with fewer parameters, averages about 77. Mistral's Ministral 3 8B, from France's best-known AI lab, averages about 71.

On the same table, Olala mini beats Europe's earlier publicly funded efforts by a wide margin, including EuroLLM and Switzerland's Apertus. That shows progress for European public research, not a lead over commercial rivals.

The hardware claim also needs care. OVHcloud says an 8 GB graphics card is enough. ActuIA noted that the full-precision model alone takes about 14 GB. The download page lists compressed versions from 6 to 14 GB, and recommends a 24 GB card for its standard high-volume serving setup.

The only named customer is Qwant, a French search engine. It uses Olala mini for the short answers at the top of results, about 3 million requests a day according to OVHcloud. OVHcloud says answers come 20% faster and running costs fell to a third of the previous model's.

Those are vendor figures, and the customer is not fully independent. ActuIA reports that Qwant's parent company, Synfonium, is 75% owned by Octave and Miroslaw Klaba, OVHcloud's founders.

Why it matters to buyers and investors

OVHcloud has until now been a supplier of servers and storage. Making its own models is a bid to keep more of the value from AI customers who would otherwise rent its machines to run Chinese or American models.

It enters with modest resources. ActuIA, citing AFP, reports that just under 100 people work on AI at the company, against more than 1,000 employees at Mistral. Klaba argues that training costs have fallen twentyfold in three years, so building models can now be profitable.

Purists will also note that OVHcloud has shared the finished model, not its training code or full datasets. That makes it open-weight rather than fully open source, the same as most rivals.

For European companies that must keep data and suppliers inside the bloc, Olala mini is a credible option for simple, high-volume jobs on cheap hardware. For anything demanding, OVHcloud's own numbers say better free options already exist. The test is whether version 2.0 arrives, and whether outside evaluators confirm the scores.