OVHcloud releases Olala mini weights, with modified runtimes needed locally
Edited and reviewed by AI Ping Editorial Team
3-line summary
- Olala mini weights and translation and summarization variants were released on November 19 under Apache 2.0.
- Local files and hosted APIs are available; standard Ollama and LM Studio support is still pending.
- Qwant uses its summarization model, with evaluation results and runtime requirements now documented.
OVHcloud released the weights for Olala mini 7A1B on November 19, moving its small language model from a November 17 announcement to a downloadable release. The Hugging Face repositories include model files and a technical report under Apache 2.0, a license allowing commercial use and modification.
Developers can run Olala locally or access a hosted API. Alongside the final assistant, the family includes a pre-trained base, an intermediate supervised checkpoint, and specialized translation and summarization models. The base and intermediate checkpoints are for research and further training; the final model serves as the assistant.
The model card specifies 6.8 billion total parameters, with about one billion active per token. Its 32,000-token context covers input and output together. The assistant supports tool calls and four settings controlling how much reasoning it performs before answering.
Local files come with runtime requirements
The GGUF distribution offers a 13.6GB BF16 file and smaller versions of roughly 6.9GB, 6.3GB and 6.0GB at eight, six and five bits. Quantization reduces the precision used to store model values. The team tested a four-bit GGUF file but withdrew it, citing poorer quality and a small additional storage saving.
On a MacBook Air M5 with 16GB of memory, the developers report 72 generated tokens per second for the six-bit file. Their short benchmark used a 512-token prompt and 128 generated tokens, with other applications open. That hardware-specific result does not establish latency for longer documents or a complete conversation.
Standard releases of Ollama, LM Studio and llama.cpp cannot yet load Olala’s architecture. The team supplies modified runtimes. Apple Silicon MLX files also need its mlx-engine fork, with an additional M1 setting to prevent a crash at the first generated token. The MLX instructions specify Python 3.11 and macOS 14 or later.
The development used JUPITER in Germany, Jean Zay in France and MeluXina in Luxembourg. The base model card describes roughly 12.8 trillion tokens from public and synthetic data, with the recipe in the technical report. Three internally generated datasets and their generation code are also released.
OVHcloud reports a 68.0 mean over twelve tests: HumanEval+ 88.3%, GPQA Diamond 50.5%. BFCL V3 scored 59.6%, outside the mean. Comparisons used recommended settings.
Tests used high reasoning effort; code and math report five-sample pass@1 averages. AI Ping has not reproduced the results.
Olala’s summarization model already powers Qwant’s flash answers above search results. A November 19 joint announcement from Qwant, Shadow, EUSP and OVHcloud says it runs on Shadow’s Blade GPU platform and handles three million requests daily. The same announcement claims one-third of the previous model’s resource use and 20% faster responses.
Blade also hosts the public playground. Developers can use OVHcloud AI Endpoints or Blade’s compatible API instead of downloading weights. The partners say their linked model catalogs let customers retain their existing provider as the entry point.
The model card describes short, single-purpose requests as its main workload. It identifies limits in stored factual knowledge, understanding outside English and major European languages, and repetition during lengthy reasoning. Korean appears in the training-language list; that inclusion alone does not establish independently verified Korean performance.
Sources
- AI Summit 2026: Olala model family — OVHcloud
- Olala v1.0 mini 7A1B — model card — OVHai LLM
- GGUF files, hardware tests and runtime requirements — OVHai LLM
- MLX files and Apple Silicon runtime requirements — OVHai LLM
- Olala v1.0 mini Technical Report — OVHai LLM
- Qwant, Shadow, EUSP et OVHcloud s’allient pour accélérer l’IA européenne souveraine — OVHcloud / Qwant / Shadow / EUSP
- Base checkpoint: training data and research scope — OVHai LLM