LATEST
AI Tools

OVHcloud Olala Mini 7A1B: A 7B Open Model That Runs on 8GB of VRAM

Updated November 17, 2026
10 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of a modern server unit in a blue-lit data center environment

Quick Answer & Key Takeaways

  • • OVHcloud announced Olala mini 7A1B on November 17, 2026, its first in-house language model: 7 billion parameters in total, with only 1 billion active per token.
  • • The model is free under the Apache 2.0 licence and runs on a GPU with as little as 8GB of memory, but OVHcloud published no benchmark scores with the announcement.
  • • Weights are promised on Hugging Face, with the formal launch at the OVHcloud AI Summit on November 19.

OVHcloud, the French cloud provider, announced its first family of in-house language models on November 17, 2026. The first member, Olala mini 7A1B, is a compact model with 7 billion parameters in total and only 1 billion active for each token. It is free under the Apache 2.0 licence, it runs on a GPU with as little as 8GB of memory, and it was trained from scratch on European public supercomputers. The catch is the one that matters most for anyone choosing a model today: OVHcloud published no benchmark scores with the announcement, and the weights were not yet public on Hugging Face at publication time. This article covers what the model is, what it costs, what it can and cannot prove yet, and how it stacks up against the small open models already available.

What Olala mini 7A1B is

Olala mini is a mixture-of-experts model. Its 7 billion parameters are spread across 256 experts per layer, and only a fraction of the network switches on for each request, which is why 1 billion active parameters is the number that matters for speed and memory. The attention design is a hybrid that combines Differential-TPA with Mamba-3, a state-space layer, in BF16 format. OVHcloud positions the model for agentic work, meaning tasks where a model calls tools and carries out multi-step jobs, rather than for simple chat alone.

The headline numbers are all OVHcloud's own, and should be read that way until independent testing exists. The company reports 160 tokens per second on an RTX A4500 workstation card, and 50 tokens per second on a standard processor with no GPU at all. The context window is 32,000 tokens today. Version 1.1, expected by the end of 2026, is promised to extend that to 128,000 tokens and improve mathematical reasoning. For comparison, Liquid AI's d1.2, launched the same week, also targets efficient on-device agent work, but with a very different architecture and a published benchmark story.

What it costs

The model itself costs nothing. Apache 2.0 is a permissive open-source licence: the weights can be downloaded, modified, redistributed and used commercially, with no licence fee and no royalty. That is the whole price story for the model. The real cost is compute, and there are two ways to pay for it.

Self-hosting is the cheap path OVHcloud is selling. An 8GB GPU is the stated minimum, which puts the model within reach of a mid-range workstation card or a modest cloud GPU instance, and CPU-only inference at a reported 50 tokens per second is usable for short, low-volume jobs. There is no subscription, no API key and no per-token meter when you run weights yourself. The trade-off is that you operate the model: you handle updates, safety filters and uptime.

The second path is managed inference, and here OVHcloud has published no Olala price. The company's AI Endpoints service sells pay-per-use inference for open-weight models, with prices starting at €0.04 per million tokens for the cheapest models in its catalogue, according to Hugging Face's write-up of the service. At November 17, 2026 market rates (€1 = $1.10), that floor is roughly $0.04, or about £0.03 and CA$0.06, per million tokens. That figure is the platform's floor, not Olala's price. Until OVHcloud lists Olala in the AI Endpoints catalogue with a rate of its own, any per-token cost for hosted Olala is unannounced. Readers planning a budget should price the self-hosted route, not assume a hosted rate.

Trained in Europe, on public machines

The training story is the part of the announcement with the most substance behind it. Olala mini was trained from scratch on three European public supercomputers, Jupiter, MeluXina and Jean Zay, using GH200, A100 and H100 GPUs. OVHcloud says the public support behind the project was equivalent to €7 million in computing hours, roughly $7.7 million, or about £5.3 million and CA$10.0 million at November 17 rates, and describes that support as decisive. The work also created OVHcloud's Agentic Lab, called OVHai LLM, a dedicated team that will keep building the family.

The release package is wider than the weights alone. OVHcloud says Olala mini ships with its pre-trained base model, a technical report and AI Act compliance documentation. The AI Act paperwork matters for European buyers: it is the kind of documentation a company needs before it can deploy a model in regulated settings. Note the wording, though. What is described is a pre-trained base model. OVHcloud has not announced an instruction-tuned or chat version, so anyone expecting a ready-made chatbot will need to fine-tune or wrap the base model themselves.

The Qwant test: 20% faster, a third of the cost

The announcement carries one live deployment. Qwant, the French search engine, has already adopted Olala mini for its flash-answer feature, the short AI answer shown above search results. OVHcloud reports answers 20% faster and running costs divided by three compared with the model Qwant previously used. Both figures are OVHcloud's own account of a deployment it is involved in, not an independent measurement, and the previous model is not named. Still, it is the only real-world evidence in the announcement, and it points at the model's intended niche: fast, cheap, short answers at scale, where a 1-billion-active-parameter model is enough and a large frontier model would be wasted money.

What the announcement does not prove

Three gaps deserve plain stating, because they decide whether this model is worth your time this week.

First, there are no benchmark scores. The press release contains no results on standard tests such as MMLU, GSM8K or HumanEval, and no comparison model is named. A small model that is fast and cheap is only useful if its answers are good enough, and nothing in the announcement lets a reader judge that. The technical report may fill the gap when it appears, but until then the quality claim rests on OVHcloud's word and one Qwant deployment.

Second, the weights were not yet public when this article was published. OVHcloud says Olala mini will be available as open source on Hugging Face under Apache 2.0, but a check of the OVHcloud organisation page on Hugging Face on November 17 showed no public models at all. The formal launch is scheduled for the OVHcloud AI Summit on November 19, so the repo may appear then. Until it does, nobody outside OVHcloud and Qwant can run the model.

Third, the 32,000-token context window is modest by late-2026 standards. Rivals in the same size class ship 128,000 tokens or more today, and Olala only reaches that figure with version 1.1 at the end of the year. For long documents or long agent sessions, that is a real limitation right now.

Olala vs the small open models

Olala enters a crowded class of small, openly licensed models. The table compares it with the three closest options, using each model's own published specifications.

ModelTotal / active parametersContext windowLicenceMinimum hardwareHosted API
Olala mini 7A1B7B / 1B32,000 (128,000 in v1.1)Apache 2.08GB GPU, per OVHcloudNone announced
Qwen3-8B8.2B / 8.2B32,768 native, 131,072 with YaRNApache 2.08GB VRAM at 4-bitYes, multiple providers
Mistral 7B v0.37.3B / 7.3B32,000Apache 2.07GB VRAM at 4-bitYes, multiple providers
Lara Mini Base7.2B / 0.6B131,072CC BY-NC-SA 4.0Not publishedNone listed

Two things stand out. Olala's 1 billion active parameters make it the cheapest to run per token in this group, on paper, because most of the network stays idle on every request. And its licence is the cleanest for commercial use: Apache 2.0, the same as Qwen3-8B and Mistral 7B v0.3, while Translated's Lara Mini Base uses a non-commercial licence that rules out business use without a separate agreement. What Olala cannot match yet is evidence. Qwen3-8B ships with published benchmark results and a year of community testing, and Mistral 7B v0.3 is a known quantity with broad tooling support. Olala has neither until the weights and the technical report land. For a closer look at the other end of the small-model spectrum, see our review of Mistral Large 4, a 1-trillion-parameter model that shows what the same lab does when it stops optimising for size.

How to try it, and what to watch on November 19

There is nothing to download yet, so the practical move is to watch the launch. The OVHcloud AI Summit on November 19 is where the company will present the model in detail, and that is the likely moment the Hugging Face repository and technical report appear. Check the OVHcloud organisation page rather than searching by model name, since the repository name is not yet confirmed. When the weights do appear, the sensible first test is a short, factual question-answering task against Qwen3-8B on the same hardware, measuring tokens per second and answer quality side by side, before committing the model to anything a user will see.

The longer roadmap is ambitious. Version 2.0 is planned to exceed 100 billion parameters, with models specialised in three areas: agentic AI for businesses, cloud DevOps, and cybersecurity, covering both defensive and offensive uses. OVHcloud also plans to use Olala to power OVHai Workspace, its encrypted office suite, where models must run locally on the workstation. Chief executive Octave Klaba framed the strategy plainly: "We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come." He also said the cost of training a model has been divided by 20 in three years. Both claims are his, and the second is the bet the whole Olala family rests on.

Common mistakes

Do not judge Olala by its 7 billion total parameters. The active count, 1 billion per token, is the figure that decides speed, memory and cost, and it is a fifth of what Qwen3-8B uses on every request. Do not assume a hosted price exists yet. OVHcloud has announced no per-token rate for Olala, so budget on self-hosting until one appears. Do not treat the Qwant figures as a benchmark. A 20% speed gain and a threefold cost cut inside one company's deployment are useful signals, not a substitute for independent scores. And do not download weights from unofficial mirrors if the Hugging Face repository is slow to appear. The official repository is the only copy with a verifiable licence and provenance, and that matters for a model you may ship commercially.

FAQ

Is Olala mini 7A1B free to use?

Yes. OVHcloud is releasing it under the Apache 2.0 licence, which permits free commercial use, modification and redistribution. You pay only for the compute to run it, whether that is your own hardware or a hosted service.

Can I run Olala mini 7A1B on my laptop?

Probably, if the laptop has a GPU with at least 8GB of memory. OVHcloud reports 160 tokens per second on an RTX A4500 and 50 tokens per second on a standard CPU with no GPU. The weights are not yet public, so no one can test those figures independently until the Hugging Face release.

When will Olala be on Hugging Face?

OVHcloud says the model will be available as open source on Hugging Face under Apache 2.0. As of November 17, the OVHcloud organisation page listed no public models. The formal launch is at the OVHcloud AI Summit on November 19, which is the most likely date for the repository to appear.

How does Olala compare with Qwen3-8B?

Both are Apache 2.0 models of about 8 billion parameters with a 32,000-token native context. The difference is efficiency: Olala activates 1 billion parameters per token against Qwen3-8B's full 8.2 billion, so it should run faster and cheaper. Qwen3-8B has published benchmarks and a year of testing; Olala has neither yet.

What are Olala v1.1 and v2.0?

Version 1.1 is expected by the end of 2026, with a 128,000-token context window and better mathematical reasoning. Version 2.0 is planned to exceed 100 billion parameters, with specialised models for business agents, cloud DevOps and cybersecurity.

Olala mini 7A1B is a serious first step from a cloud provider that has decided to stop depending on other labs for models, and its 1-billion-active-parameter design is an interesting answer to the cost of running AI. But a model with no published benchmarks and no public weights is a promise, not a product. Check the Hugging Face repository after November 19, run it against Qwen3-8B on your own hardware, and let the numbers decide. For another recent example of a small model built for on-device work, read our coverage of Edge0-145B, a 145B model that runs on a laptop.

10 min read
140 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more