OVHcloud releases open Olala mini 7A1B: 7 billion parameters, runs on an 8 GB GPU
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Open weights under Apache 2.0 from November 19 will make it possible to run the model on a single graphics card with as little as 8 GB — without calling on an external cloud.
OVHcloud on November 17 unveiled Olala, its first in-house family of large language models; the first to be released is the compact Olala mini 7A1B. The model is sparse: 7 billion parameters, of which 1 billion are active, 256 experts per layer, hybrid attention on Differential-TPA and Mamba-3, a 32,000-token context, BF16 format. According to the company, it runs on a GPU with 8 GB of memory (160 tokens per second on an RTX A4500) and on a CPU (50 tokens per second). Training took place on the European supercomputers Jupiter, MeluXina and Jean Zay — the equivalent of €7 million in compute time. The weights will be published on Hugging Face under the Apache 2.0 licence from November 19, together with the base pre-trained model, a technical report and AI Act documentation. The Qwant search engine already uses the model for quick answers: according to the company, answers in "Réponse Flash" became 20% faster at three times lower cost. Version 1.1 with a 128,000-token context is promised by the end of the year, and version 2.0 with more than 100 billion parameters later on.
Sources 3