OVHcloud has launched Olala mini 7A1B, a seven-billion-parameter language model that the company says can run on a standard CPU without a GPU. Announced on November 19, 2026, it kicks off OVHcloud's first language-model family, with weights available on Hugging Face. The GPU, for once, isn't on the compulsory shopping list. SourceSource
According to its AI unit, OVHai LLM, the model range was designed and built from scratch in France. The company says Olala mini was trained on three public European supercomputers: Jupiter, MeluXina and Jean Zay. Source
OVHai says only a fraction of the model is used for each request, and its comparison table lists roughly one billion active parameters. That's the efficiency pitch behind the CPU claim—not a promise that every laptop will deliver the same performance. Source
No GPU doesn't mean no setup
In its November 20 coverage, Underclassic reported, citing the company, that standard llama.cpp, Ollama and LM Studio could not yet load Olala mini's new architecture. TeksEdge's November 20 post likewise said the GGUF version required OVHcloud's custom llama.cpp fork—a modified version of the runtime. SourceSource
Five variants, with small jobs in mind
OVHai lists five models in the range: the generalist mini 7A1B, Translate, Summarise, Base and SFT. The company describes Translate as supporting 18 languages and Summarise as producing structured summaries. Base is the pre-trained model, while SFT is the supervised fine-tuned checkpoint used before reinforcement learning. The launch page also links to a technical report. Source
OVHai's deployment examples cover database queries, extracting information, producing structured output and calling tools. In an engineering example, OVHai says the model completes shorter files within a workflow that divides the task and checks each result. OVHai recommends that arrangement: “the workflow runs the process, Olala mini does the steps, and code checks the numbers.” The calculator still gets to keep its job. Source
Comments