OVH’s Olala mini 7A1B model uses 1B active parameters out of a 7B total and was trained on 13T tokens. The model utilizes a custom architecture mixing Mamba-3 with grouped differential attention layers and achieves 220-380 tokens/sec on Blade B300 hardware. Because it activates only 1B parameters per request, the model can run agentic tasks without a GPU.
The company will release version 1.1 in six weeks, increasing the context window to 128K. OVH plans a version 2.0 with more than 100B parameters designed for devops, cybersecurity, and agentic workflows.
