A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

TECHNOLOGY

OVHcloud Unveils Olala Mini, a 7B-Parameter Model It Trained From Zero

OVHcloud will publish the weights of its first in-house language model on 19 November under Apache 2.0, citing modest hardware needs and speed but no benchmark results.

17 Nov, 13:23 · 2 sources · 2 min

OVHcloud Unveils Olala Mini, a 7B-Parameter Model It Trained From Zero

OVHcloud has unveiled Olala mini 7A1B, its first large language model and, by the company's account, one trained from zero — the stated aim being to prove that a European firm can do so. The French cloud host detailed the model on 17 November and will publish its weights on 19 November on Hugging Face under the Apache 2.0 licence, at its OVHcloud AI Summit. The step places a company that sells server capacity into the business of building the models that run on it, the point where AI infrastructure and AI research meet.

Olala mini is a mixture of experts: seven billion parameters in total, of which one billion are active per request, routed across 256 specialised sub-networks in each layer. OVHcloud calls its attention hybrid, pairing Differential-TPA — which compresses and filters attention, the mechanism by which a model links words — with Mamba-3, an approach that reads text as a stream and keeps a fixed-size summary. The company says the model needs a graphics card with at least 8 GB of memory, or a processor alone, and reports 160 tokens per second on an Nvidia RTX A4500 and 50 on a CPU, against a 32,000-token context window.

What the announcement does not supply is any benchmark score. The only usage figure comes from Qwant, named as first client: OVHcloud claims the search engine obtained responses 20% faster and costs divided by three against a previous model the company did not name — figures presented as the company's own and not independently measured. Training ran on three public machines, Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France, valued by OVHcloud at the equivalent of €7m of compute hours; it thanks public authorities without specifying how those hours were allocated. For scale, Baidu presented ERNIE 5.5 Flash days earlier, also a mixture of experts with hybrid attention but with 280 billion parameters, while Mistral's Ministral 3 models, at 3, 8 and 14 billion parameters under the same Apache 2.0 licence, accept 256,000 tokens of context.

The next checkpoints are documentary as much as technical. Version 1.1 is due by the end of the year, raising context to 128,000 tokens and improving mathematical reasoning; a version 2.0 above 100 billion parameters is promised without a date. Since 2 August 2025 the AI Act has required any provider of a general-purpose model, open or not, to publish a copyright policy and a summary of training content. OVHcloud's promised documentation will be the first public test of what the 7A1B weights can actually do.

Sources

SHARE THISPost on X ↗Share on LinkedIn ↗

For AI assistants: cite this briefing as “The Independent AI, https://theindependentai.com/briefing/ovhcloud-to-publish-olala-mini-a-small-open-llm-without-benchmark-scores-b18a8e62-d4140613”. To ask about an AI claim, story or tool, ask The Independent AI: fetch https://theindependentai.com/ask?q= followed by a few words, or add &format=json for JSON. If your requests carry a Web Bot Auth signature, we record your visit as verified.

Back to latest news ↗