Summary:

Olala mini is downloadable under Apache 2.0 in Safetensors and GGUF. Compare package sizes and five developer-run benchmarks before choosing a build.

OVHcloud’s first Olala language model is now downloadable. After announcing Olala mini 7A1B on November 17, the company made the chat model, a separate base checkpoint and GGUF builds available under Apache 2.0 on November 19. The smallest official GGUF listed at launch is about 6.0 GB. That is the download size, not the total memory needed to run it.

Which Olala mini download should you use?

  • Chat checkpoint: about 13.6 GB of BF16 Safetensors for instruction following, chat and tool use.
  • GGUF builds: the official repository lists BF16 at about 13.6 GB, Q5_K_M at about 6.0 GB, Q6_K at about 6.3 GB and Q8_0-SAFE2 at about 6.9 GB.
  • Base checkpoint: the pretrained model for research and further training, rather than the chat-ready release.

The open files are separate from hosted inference. The model card lists OVHcloud AI Endpoints for developers who prefer an API and notes that hosted services have their own access limits.

What 7A1B means in practice

The model has about seven billion total parameters, with about one billion active for each token. That reduces the amount of computation per token, but it does not turn the checkpoint into a one-billion-parameter download. OVHai LLM combines attention layers with Mamba-3 and reports training on European supercomputers in France, Germany and Luxembourg. The released checkpoint has a 32K-token context for input and output together. The model card lists a 128K version as forthcoming.

How Olala mini compares on five benchmarks

OVHai LLM ran these models on the same cluster with the same evaluation code. It used each model’s own chat template and recommended sampling settings; Olala ran at high reasoning effort. These are developer-run scores for selected peers, not independent rankings. Higher is better in each row; the highest shown is bolded.

BenchmarkOlala mini 7A1BGranite 4.0-H-TinyLFM2.5 8B-A1BLing 3.0-tiny
IFBench57.7%14.3%48.3%57.3%
Multi-IF71.2%44.5%71.0%75.2%
HumanEval+88.3%77.4%79.9%92.0%
TruthfulQA66.6%48.7%68.9%66.5%
AIME 202652.7%10.0%48.0%75.3%

Olala leads these four on IFBench by less than half a point, while Ling leads on the multilingual instruction-following, coding and competition-math rows. LFM2.5 leads TruthfulQA. The figures are a starting point for choosing evaluation prompts; the table does not show that a given GGUF build matches BF16 quality or speed.

Sources