Hugging Face Hosts OVHaiLLM's Olala-v1.0-mini-7A1B Model for Causal Language Modeling
Hugging Face is hosting the 'OVHaiLLM/Olala-v1.0-mini-7A1B' model, a compact reasoning model designed for chat, instruction following, and tool use. Developed by OVHai LLM (OVHcloud group) in France, this model utilizes a hybrid Mamba-3/attention mixture-of-experts architecture. It was trained on three European supercomputers—JUPITER, Jean Zay, and MeluXina—using 12.8 trillion tokens from public datasets and synthetic data. The model is released with open weights under the Apache 2.0 license. It is efficient, with approximately 1 billion active parameters per token, and can run on a laptop. The Olala-v1.0-mini-7A1B model is designed to reason before providing answers, offering four levels of reasoning effort, and has been trained with reinforcement learning for tool calling and multi-turn searches. It supports various tasks including general chat, instruction following, mathematics, science, code, and multilingual applications. The model can be deployed on NVIDIA GPUs or CPUs and integrates with the Transf...