Osegi
AI·2 min read

OVHcloud Olala mini: a free AI model that runs in 8 GB of RAM

OVHcloud has released Olala mini on Hugging Face, with benchmarks this time. It follows instructions well, struggles with science, and won't yet load in Ollama.

A small laptop on a desk showing a chatbot interface, with server racks blurred in the background.
Image: Frandroid

Two days after unveiling Olala mini 7A1B without a single benchmark score, OVHcloud has now put the model online and shown its numbers. The French host released it on Hugging Face on Thursday, at its AI Summit in Paris, with a technical report and a comparison against 18 other small models.

It is not one model but five: the final chat and reasoning version, the raw base model, an intermediate version, and two specialised ones for summarising documents and for translation. All use the Apache 2.0 licence, which allows commercial use. OVHcloud also says it is publishing its training data recipe and training tools, which is rare. Most "open" models only share the weights. Olala was trained on 12,800 billion tokens in English and 36 other languages.

How good is Olala mini?

Across twelve tests covering instruction following, science, code and maths, Olala averages 68 out of 100, which puts it seventh. It beats Ministral 3 3B but trails Ministral 3 8B. It does far better than the European models in the table, such as EuroLLM 9B and Apertus 8B, which score under 32.

It has 7 billion parameters but uses only 1 billion for each word it writes, so it computes less and runs faster. Its strength is instructions: 57.7% on IFBench, double Ministral 3 8B. Its weakness is science: 50.5% on GPQA Diamond against 75.8% for Qwen3.5 4B, which beats it on nine of twelve tests. These are OVHcloud's own figures, and no independent group has reproduced them.

Can you run Olala mini on your own PC?

Yes, but not with the easy tools. Ollama, LM Studio and the official llama.cpp can't load it yet. You need modified versions published by OVHcloud, through the command line. The smallest compressed file is 6 GB, and the uncompressed model needs 13.6 GB. On a MacBook Air M5 with 16 GB, the 6-bit version produces about 70 tokens per second, according to the company. To just try it, there is a chatbot at olala.ovh.com.

The first customer is Qwant, whose "flash" answers now use the summarising version, handling 3 million queries a day with a third of the resources and answers 20% faster, says OVHcloud. Qwant is owned by a company co-held by the Klaba family, and the model runs on Shadow GPUs, so the client is not quite external.

Version 1.1 is due by the end of the year, with a context window growing from 32,000 to 128,000 tokens. Version 2.0 will pass 100 billion parameters.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:ovhcloudolala miniopen source aimistralhugging faceqwant

More from this day

See all →

More in: AI

See all →