Olala — local AI model by OVHcloud
Written by Jakub Rusinowski · Last updated
OVHcloud's first family of language models, trained from scratch on three European supercomputers (JUPITER, Jean Zay and MeluXina). Olala mini 7A1B is a mixture-of-experts with 6.8B parameters in total and about 1B active per token, built on a hybrid of 29 Mamba-3 layers and 7 attention layers, with a 32K context, four reasoning-effort levels and tool calling. OVHcloud publishes the weights, the base model and a technical report under Apache 2.0. Today every local route uses one of OVHcloud's forks: stock llama.cpp, vLLM, mlx-lm, Ollama and LM Studio cannot load it yet.
Variants
The smallest Olala variant needs about 7 GB of VRAM at Q5_K_M — quantized weights plus framework overhead, before any KV cache.
| Model | VRAM at Q4 | VRAM | Context | Run it |
|---|---|---|---|---|
| Olala mini 7A1B → 6.8B (~1B active) | ~6.8 GB | 32,768 | cloud-hosted tag |
Memory is quantized weights plus overhead at Q4_K_M, from the same engine as the GPU & VRAM checker.
How to run Olala locally
Install Ollama, then pull the tag.
No Ollama or LM Studio route: OVHcloud's official GGUF and MLX builds need its forks of llama.cpp and mlx-engine, and server use needs its vLLM image (ovhaillm/olala-vllm). Stock llama.cpp, vLLM and mlx-lm cannot load the olala architecture yet.
Pick a size above for its own VRAM figure, speed estimate and install command.
Licence
Commercial use permitted. No usage restrictions beyond attribution.
Applies to: Olala mini 7A1BRecommended GPU
The cheapest catalogued GPU that runs Olala locally (min 7 GB VRAM) is the Intel Arc B570 (10 GB).
Olala — frequently asked questions
How much VRAM does Olala need?
Olala needs about 7 GB VRAM at Q5_K_M quantization for its smallest variant. Variants: Olala mini 7A1B (7 GB, Q5_K_M). On Apple Silicon, unified memory counts toward this requirement.
Can I run Olala on an RTX 4090 (24 GB)?
Yes — Olala runs on an RTX 4090 (24 GB) and other 24 GB cards such as the RTX 3090. Smaller variants also fit comfortably on 8–16 GB GPUs at Q5_K_M.
What quantization should I use for Olala?
Olala is published only pre-quantised, so there is no Q4_K_M or Q8_0 build to choose between: each variant ships in the format listed beside it above (Q5_K_M). Use the vendor's runtime for that format, as the variant pages describe.
How do I run Olala with Ollama?
Olala has no Ollama tag yet. No Ollama or LM Studio route: OVHcloud's official GGUF and MLX builds need its forks of llama.cpp and mlx-engine, and server use needs its vLLM image (ovhaillm/olala-vllm). Stock llama.cpp, vLLM and mlx-lm cannot load the olala architecture yet.