OVH Enables Download of Its AI Tailored for 8GB RAM, and Its Scores Place It Against Mistral

OVH Enables Download of Its AI Tailored for 8GB RAM, and Its Scores Place It Against Mistral
3 min read

OVHcloud has made Olala mini 7A1B, its first large language model, freely available, on Thursday, November 19 in Paris during its AI Summit. Presented two days earlier without any test results, it is now accompanied by a technical report. According to the hosting provider itself, it is a mid-tier model, very good at following instructions, but less proficient in mathematics and science.

OVHcloud is not publishing one model, but five. The final version, designed for chatting, reasoning, and using tools, is accompanied by the base model, raw, without tuning for conversation, an intermediate version, and two specialized variants, one for document summarization, the other for translation. All are under the Apache 2.0 license, which permits commercial use without compensation.

The company also announces publishing the recipe for its training data as well as its training tools, a rare practice, as most so-called ‘open’ models are limited to the weights, i.e., the model itself. Olala was trained on 12.8 trillion tokens, text fragments, in English and 36 other languages, primarily European.

Seventh out of nineteen

OVHcloud compared Olala to 18 models of comparable size, across twelve tests covering instruction following, scientific knowledge, coding, and mathematics. With an average score of 68 out of 100, the model ranks seventh. These scores are the company’s own, measured with its own settings, and no independent organization has reproduced them yet.

Against Mistral’s models, the result is mixed: Olala outperforms Ministral 3 3B, but remains behind Ministral 3 8B. It achieves significantly better results than other European models on the chart, such as EuroLLM 9B or Apertus 8B, whose average is below 32. Out of its 7 billion parameters, the settings learned during training, it only activates one billion for each word produced, compared to 4 billion for Qwen3.5 4B and 8 billion for Ministral 3 8B. It therefore computes less and is faster, but Qwen3.5 4B performs better on nine out of twelve tests.

OVHcloud itself acknowledges that its model is ‘stronger at following instructions, weaker at competition-level mathematics.’ On IFBench, which checks compliance with specific constraints, like a word count or a required format, Olala scores 57.7%, double that of Ministral 3 8B. On GPQA Diamond, a university-level science test, it reaches 50.5%, compared to 75.8% for Qwen3.5 4B. Olala is designed for short, repetitive tasks, like summarizing, classifying, extracting information, or calling the right tool. It was trained to chain up to 75 tool calls in a row.

On your PC, not yet with two clicks

OVHcloud promised a model that ‘fits on a PC without a dedicated graphics card.’ This is not yet possible with mainstream tools: Ollama, LM Studio, and the official version of llama.cpp cannot load it yet. One must use modified versions of these software, published by OVHcloud, and go through the command line.

Once installed, the model remains lightweight. The smallest compressed version weighs 6 GB, and OVHcloud has not published a more compressed version, due to insufficient gain. On a MacBook Air M5 with 16 GB of memory, the 6-bit version generates about 70 tokens per second, according to the company. The uncompressed version requires 13.6 GB.

Olala already has one client, the search engine Qwant. Its ‘flash’ answers, displayed at the top of results, now rely on the version specialized in summarization. It processes 3 million queries per day, with three times fewer resources than the previous model, itself from Mistral, and answers 20% faster, according to OVHcloud. Qwant has been owned since 2023 by a company jointly held by the Klaba family and the Caisse des dépôts, and the model runs on Shadow’s GPUs, acquired in 2021 by a company of Octave Klaba. The three companies announced an alliance this Thursday for an ‘end-to-end sovereign AI.’ Nothing, however, was said about the quality of the summaries.

A version 1.1 is expected by the end of the year. Its context window, i.e., the amount of text the model keeps in memory, will increase from 32,000 to 128,000 tokens. Version 2.0 will exceed 100 billion parameters. Speaking to the press, Octave Klaba gave a range of 100 to 300 billion for early 2027, and ultimately aims for a model with 3 trillion parameters for cybersecurity and development. Developed by a team of about a hundred people, Olala mini stands out primarily for its complete openness, including training data.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x