OVHcloud makes Olala mini weights available for local deployment

OVHcloud’s Olala mini weights are now publicly available on Hugging Face, moving beyond the company’s November 17 announcement of a planned release. The official repository, checked on November 19, includes a 13.6 GB model file, inference code and a technical report.
The model card specifies 6.8 billion total parameters, with approximately one billion active per token, and an Apache 2.0 license. Its hybrid design combines Mamba-3 and attention layers. The current context window is 32,000 tokens; a 128,000-token version remains forthcoming.
The text-only model supports chat, reasoning and tool calls, with four reasoning-effort settings. OVHcloud also lists hosted access through its AI Endpoints service and provides separate quantized variants for local use.
Local compatibility requires care: the developer says stock Ollama, LM Studio, llama.cpp, vLLM and mlx-lm cannot yet load this architecture. Its documented local routes rely on customized software forks.
Downloadable weights let developers evaluate and adapt the model directly, but availability should not be confused with plug-and-play support across popular desktop AI applications.
Sources and citations
Revision history
Published .