Olala mini 7A1B / Olala
Intro
Olala is the first family of large language models announced by the French cloud company OVHcloud on 17 November 2026. The first public member is Olala mini 7A1B. The official English newsroom post is titled “OVH announces Olala mini 7A1B” and dated Tue, 11/17/2026 - 12:00. It opens: Roubaix – 17 november 2026 – OVH today announces Olala, its first family of LLMs. The company separates this announcement from the official launch at the OVHcloud AI Summit on 19 November at the Carrousel du Louvre in Paris: the name and specs were published on the 17th; the 19th is when it plans to present the work in more detail and release the weights.
According to the newsroom post, Olala mini is trained from scratch as a mixture-of-experts model: about 7 billion parameters in total, 256 experts per layer, about 1 billion parameters active per request, hybrid attention that combines Differential-TPA and Mamba-3, BF16 precision, and a 32,000-token context window. OVHcloud says it runs on a GPU with as little as 8 GB of memory, at about 160 tokens per second on an RTX A4500, and at about 50 tokens per second on a standard CPU with no GPU. Those are official speed figures; the public announcement does not include quality scores.
Training used three European public supercomputers, Jupiter, MeluXina and Jean Zay (the post mentions GH200, A100 and H100 GPUs). OVHcloud puts the public compute support at the equivalent of €7 million in computing hours and says this also led to its Agentic Lab, OVHai LLM. The post also says French search engine Qwant already uses Olala mini for its flash answer feature, with replies 20% faster and running costs cut to one third of the previous model — official claims. Do not confuse this with older personal or fine-tune repositories that only share a similar name.
Advertisement
Explain
Olala mini is positioned as a compact, European-trained model that can run on a workstation, not as a flagship dense model. The 7A1B suffix matches the published spec: 7B total parameters and about 1B active. The company says only a fraction of the experts is used for each request, which is meant to cut memory use and raise speed. The planned licence is Apache 2.0, and the release package is described as the weights, the pre-trained base model, a technical report, and AI Act compliance documentation. As of this scan, no official Olala weights repository had shown up on Hugging Face, which matches the official timetable of publishing them at the 19 November launch.
The same newsroom post sketches the roadmap: v1.1, expected before the end of the year, would extend the context window to 128,000 tokens and improve mathematical reasoning; v2.0 is planned to exceed 100 billion parameters and to specialise in three areas — agentic AI for businesses (eventually powering a local OVHai Workspace), cloud DevOps, and cybersecurity. CEO Octave Klaba is quoted saying OVH cannot simply rely on other labs to keep sharing open-source LLMs, that training cost has fallen by a factor of about 20 in three years, and that it is now possible to train LLMs and make that work pay. Those are official statements; independent quality comparisons will have to wait for the weights and the technical report on the 19th.
Inside this scan window, a Chinese recap post (for example LenardReads at 2026-11-17T14:31:17Z) also mentioned the Olala launch and the Qwant cost claim. Engagement on that recap was low; this entry follows the official newsroom post.
Usage
On the announcement day (17 November) the public artefact is the official write-up, not a downloadable checkpoint. Read the newsroom post: OVH announces Olala mini 7A1B. OVHcloud says the official launch is at the OVHcloud AI Summit on 19 November 2026, with Apache 2.0 weights, the pre-trained base, a technical report and AI Act documentation to appear on Hugging Face then. Wait for that official repository; do not treat an older lookalike project as Olala mini.
The official runtime guidance is a GPU with at least 8 GB of memory, or a standard CPU at lower throughput. The examples in the post are about 160 tokens per second on an RTX A4500 and about 50 tokens per second on CPU; use the forthcoming technical report for the actual setup. The in-production example cited by OVHcloud is Qwant’s flash answer feature, which is an internal substitution, not a public API guide. The press contact at the end of the post is [email protected].