Skip to content
Chain AI Wire
AI Models

OVHcloud Opens the Weights on Olala Mini 7A1B: A 7B Model That Runs on a Laptop

OVHcloud has released the open weights for Olala mini 7A1B, a 7-billion-parameter reasoning model that activates only about 1 billion parameters per token and runs on a laptop, the first model from Europe's largest cloud provider.

OVHcloud Olala mini 7A1B open-weight AI model running on a laptop
OVHcloud has released the open weights for Olala mini 7A1B, a small reasoning model that runs on a laptop.

At a glance: OVHcloud has released the open weights for Olala mini 7A1B, the first model from its in-house AI lab, at the company’s AI Summit on 19 November. The model carries 6.8 billion total parameters but activates only about 1 billion per token, reads 32,000 tokens of context, and is built to run on a laptop or an 8 GB graphics card rather than a data centre. The weights are free to download under Apache 2.0, alongside a technical report and the EU AI Act paperwork. The benchmark figures in this article are OVHcloud’s own, published on the model’s Hugging Face page, and no independent lab has had time to reproduce them.

The launch had been trailed since 17 November, when OVHcloud founder Octave Klaba introduced Olala in an interview and the company put up a Hugging Face page with the announcement. The weights themselves landed on 19 November, together with the base model, the fine-tuned checkpoint, two specialist variants for summarisation and translation, and GGUF and MLX builds for local use. OVHcloud published the Olala mini 7A1B model card on Hugging Face on 19 November. For a cloud provider better known for renting out servers than training models, it is a first step into a crowded field, and the pitch is not raw power but a model small enough to run anywhere.

What Olala mini actually is

Olala mini is a mixture-of-experts model, and the parameter count needs the usual unpacking. The headline figure is 6.8 billion total parameters, but only around 1 billion of them fire for any given token. Each of the model’s 36 layers pairs either a Mamba-3 block or an attention layer with a mixture-of-experts block holding 256 routed experts, six of which are active per token, plus one shared expert. Twenty-nine of the layers use Mamba-3, a state-space design that keeps memory use low, and seven use attention, one in five. Only seven layers keep a key-value cache at all, which is a large part of why the model fits on modest hardware.

Training ran to 12.8 trillion tokens: 10.4 trillion in the main pre-training run, then 2.38 trillion of annealing where mathematics became the largest share. The model was pre-trained on English and 36 other languages, mostly European, plus Arabic, Chinese, Japanese and Korean. The compute came from three European public supercomputers, JUPITER in Germany, Jean Zay in France and MeluXina in Luxembourg, with the base model pre-trained on 1,024 GH200 GPUs on JUPITER. OVHcloud puts the whole training run at about 1.16 GWh of energy and roughly 40 tonnes of CO2 equivalent, covered by renewable supply contracts.

The model thinks before it answers, with four reasoning-effort levels from disabled to high, and it was trained with reinforcement learning to call tools and to search over many turns, up to 75 turns in OVHcloud’s agentic training. That makes it a model aimed at assistants and agents rather than raw question answering, and the company is candid about the trade-off: with about a billion active parameters, it stores less factual knowledge than larger models and can state wrong facts with confidence. Its own guidance is to give it tools and sources rather than ask it to answer from memory.

Benchmarks: OVHcloud’s own numbers

OVHcloud ran its own evaluation across 12 benchmarks, scoring the model and 18 rivals on the same cluster with the same code. Olala mini averages 68.0 across those twelve, 15.2 points above its own supervised fine-tuning checkpoint. Its strongest results are on code and instruction following: 88.3 on HumanEval+, 75.1 on MBPP+, 57.7 on IFBench and 71.2 on Multi-IF. On knowledge tests it scores 50.5 on GPQA Diamond and 66.6 on TruthfulQA. On competition maths, its weakest area by the company’s own admission, it reaches 60.7 on AIME 2024 and 61.3 on AIME 2025. Tool calling lands at 59.6 on BFCL V3. Every one of those figures comes from OVHcloud, so treat them as the vendor’s case for the model until outside testing weighs in, and note that several rivals in the same table beat it on individual tests.

Speed is where OVHcloud makes its boldest claims. The company says the model reaches 160 tokens per second on an RTX A4500 workstation card and about 50 tokens per second on a standard CPU, and that it runs on a GPU with as little as 8 GB of memory. French search engine Qwant already uses it for its flash answers, and OVHcloud reports 20% faster answers at one third of the previous cost. Those are the vendor’s numbers, and the real test starts now that anyone can download the weights and run them.

ModelTotal parametersActive per tokenContextAvailability
OVHcloud Olala mini 7A1B6.8B~1B32KOpen weights, Apache 2.0
Tencent Hy4 Flash220B~12B1MOpen weights, Tencent Hy License
Alibaba Qwen4-Max2.4TNot disclosed1MAPI
Google Gemini 4 NeonNot disclosedNot disclosed1MAPI
OpenAI GPT-6.2 AstraNot disclosedNot disclosedNot disclosedAPI

How to run it, and the catch

The easiest route is hosted. OVHcloud serves the model through its AI Endpoints, an OpenAI-compatible API hosted in Europe, and the same service is available through Hugging Face Inference Providers. A third option, Blade, offers an OpenAI-compatible API on European GPUs with a free tier. For running it yourself, OVHcloud publishes GGUF files for llama.cpp on CPUs and Apple Silicon, an MLX build for Macs, a vLLM image for GPU servers, and Transformers code for research.

Here is the catch, and OVHcloud states it plainly on the model card. Upstream Ollama, LM Studio, llama.cpp, vLLM and mlx-lm cannot load Olala 1.0 mini yet, because none of them supports the model’s hybrid Mamba-3 architecture. Each route needs one of OVHcloud’s forks for now. That is real friction for a model whose whole pitch is that it runs anywhere, and it will matter how quickly the mainstream tools add support. The company also warns that generic quantisation recipes produce NaN outputs on realistic inputs, so users should stick to the GGUF and MLX files it publishes.

The European angle

The launch is as much about sovereignty as about the model itself. Olala is designed in France, trained on European supercomputers, released with open weights under Apache 2.0, and shipped with the documentation the EU AI Act requires, including a public summary of its training content. OVHcloud is the largest cloud provider headquartered in Europe, and the model arrives as European governments and companies look for AI they can run on their own infrastructure without sending data to US providers. A 128,000-token context version, v1.1, is promised by the end of the year, and a v2.0 with more than 100 billion parameters is planned after that. The timing is worth noting. Tencent Hy4 Flash opened its weights two days earlier, and the open-weight race is now moving in weekly steps rather than quarterly ones. Olala is the smallest model in this wave, and also the most openly European, a combination that gives it a clear audience even if it never tops a benchmark table.

That audience is the point. K2-Horizon opened a 375-billion-parameter fleet a day before Hy4 Flash, and the frontier labs keep shipping bigger models every week. Olala goes the other way: small enough to run on hardware people already own, open enough to use commercially, and European enough to satisfy procurement rules that US models cannot. Whether that is enough to matter will be decided by the independent tests now under way.

What to watch

  • Independent testing: whether outside labs reproduce OVHcloud’s benchmark average and the speed claims on ordinary hardware.
  • Tool support: how quickly Ollama, LM Studio and upstream vLLM add the Olala architecture, removing the need for forks.
  • The 128K version: whether v1.1 lands by the end of the year as promised, and what it does to the model’s weakest area, maths.
  • Qwant in production: whether more European companies follow Qwant onto the model for search and assistant work.
  • The bigger model: what v2.0 looks like at more than 100 billion parameters, and whether it stays open.

What is OVHcloud Olala mini 7A1B?

OVHcloud’s first in-house language model, a small reasoning model with 6.8 billion total parameters and about 1 billion active per token. It was released with open weights under Apache 2.0 on 19 November 2026, and is designed for chat, instruction following and tool use.

Can I run Olala mini on my own computer?

Yes. OVHcloud says it runs on a GPU with as little as 8 GB of memory, or on a CPU at about 50 tokens per second, and it publishes GGUF files for llama.cpp and an MLX build for Macs. The catch is that upstream Ollama, LM Studio, llama.cpp, vLLM and mlx-lm cannot load the model yet, so you need one of OVHcloud’s forks for now.

How much does Olala mini cost?

The weights are free to download under Apache 2.0, including for commercial use. Hosted access is available through OVHcloud AI Endpoints, Hugging Face Inference Providers or Blade, all OpenAI-compatible APIs.

How does Olala mini score on benchmarks?

On OVHcloud’s own 12-benchmark evaluation it averages 68.0. Its strongest results are 88.3 on HumanEval+ and 75.1 on MBPP+ for code, and 71.2 on Multi-IF for instruction following. It scores 50.5 on GPQA Diamond and 60.7 on AIME 2024. All of these figures are OVHcloud’s own reporting and have not been independently reproduced yet.

What architecture does Olala mini use?

A hybrid mixture-of-experts design: 36 layers, 29 using Mamba-3 and 7 using attention, each followed by a mixture-of-experts block with 256 routed experts, six active per token, plus one shared expert. The context window is 32,000 tokens, with a 128,000-token version promised by the end of 2026.

David
Written by

David

David is a senior technology and digital assets contributor at ChainAIWire, covering breaking developments across generative AI, autonomous agent ecosystems, and blockchain infrastructure. He tracks how emerging machine learning models intersect with decentralized finance, regulatory policy, and modern market structure. When he isn't analyzing protocol updates and macro liquidity shifts, he closely monitors open-source developer communities.

More articles by this author →

Leave a Comment