OVHcloud unveils Olala mini, an open LLM for 8GB GPUs
OVHcloud says its first Olala model can run on GPUs with as little as 8GB of memory and will be released under Apache 2.0.
OVHcloud has announced Olala mini 7A1B, the first model in a new family of large language models that the European cloud provider says it built from scratch. The announcement arrives ahead of the company’s AI Summit on 19 November, where it plans to present its wider cloud and artificial-intelligence work. 1
A smaller model with a specialist design
Olala mini uses a 7-billion-parameter architecture, with 1 billion active parameters distributed across 256 experts per layer. OVHcloud says it combines Differential-TPA and Mamba-3 attention in BF16 format, using only part of the model for each request. The company claims that approach lets Olala mini run on a GPU with as little as 8GB of memory.
OVHcloud reports speeds of up to 160 tokens per second on an RTX A4500, and 50 tokens per second on a standard processor without a GPU. Those are vendor-reported figures rather than independent benchmark results, so they should be treated as an initial performance claim rather than a settled comparison. The model currently has a 32,000-token context window. 1
Open release, European training
The company says Olala mini will be released on Hugging Face under the Apache 2.0 licence, alongside its weights, pre-trained base model, technical report and AI Act compliance documentation. OVHcloud says it trained the model on the Jupiter, MeluXina and Jean Zay supercomputers, with public support equivalent to €7 million in computing hours.
Qwant is already using Olala mini for its flash-answer feature, according to OVHcloud. The company says answers are 20% faster and operating costs are one-third of the previous model’s, although those figures also come from the announcement rather than an independently reproduced test.
What happens next
OVHcloud says version 1.1 is expected by the end of 2026 with a 128,000-token context window and improved mathematical reasoning. Its longer-term v2.0 plan targets models above 100 billion parameters for business agents, cloud DevOps and cybersecurity. Those plans are a roadmap, not features available in today’s release.
Our opinion
Olala mini is interesting because it targets the practical middle ground between headline-grabbing frontier models and tiny local experiments. An 8GB GPU threshold could make a capable open model easier to test on ordinary workstations, but the useful proof will come when the weights, technical report and reproducible benchmarks are available. For now, OVHcloud has made a credible announcement; it has not yet supplied enough independent evidence to establish how Olala mini compares with similarly sized open models. 1