OVH Launches Its First AI Model Family, Olala, Open-Sourcing Olala Mini 7A1B
1 day ago / Read about 0 minute
Author:小编   

European cloud services company OVHcloud has released its first AI model family, Olala. The first member, Olala Mini 7A1B, adopts a hybrid attention mechanism combining Differential-TPA and Mamba-3, with a total of 7 billion parameters and 1 billion activated parameters, uses the BF16 data format, and will be open-sourced under the Apache 2.0 license. The model can run on GPUs with 8GB of VRAM, achieving a token throughput of 160 TPS on an RTX A4500, and can also reach 50 TPS on a pure CPU platform. The model relies on three European public supercomputers and was trained using NVIDIA GH200, A100, and H100, with a total cost of 7 million euros. The initial version has a context length of 32K, and version 1.1, expected to be released by the end of the year, will expand to a 128K context and optimize mathematical capabilities. The subsequent Olala 2.0 will have a parameter scale exceeding 100B.