What OVHcloud released
French cloud provider OVHcloud announced Olala on 17 November 2026. Olala is the company's first family of language models, and Olala mini 7A1B is its first member. The weights went up on Hugging Face on 19 November under Apache 2.0, according to the model card. The card describes a text-only reasoning model for chat, instruction following and tool use. OVHcloud also published a pre-trained base checkpoint, a supervised fine-tuning (SFT) checkpoint, and separate summarisation and translation variants, along with GGUF and MLX quantised builds.
Apache 2.0 places no territorial or user-count limits on the weights, so a New Zealand team can ship Olala in a commercial product without negotiating a separate licence. OVHcloud's CEO, Octave Klaba, gave a strategic reason for training an in-house model: "We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come," he said in the press release. The release also includes EU AI Act compliance documentation, as IT Brief reported.
How the architecture works
Olala mixes two kinds of layer. According to the model card, 29 of its 36 layers are Mamba-3 layers and the remaining 7 are attention layers. Mamba is a state space model design. Instead of keeping a key-value cache that grows with every token, a Mamba layer carries a fixed-size running state through the sequence. Long prompts therefore cost less memory than they would in a pure transformer, and the few attention layers keep the model's ability to look back at exact earlier tokens. The press release names the attention variant as Differential-TPA.
Every layer feeds into a mixture-of-experts (MoE) block. An MoE block holds many small feed-forward networks, called experts, and a router picks a few of them for each token. Olala has 256 routed experts per layer, activates 6 of them for each token, and adds 1 shared expert. The model holds 6.8B parameters in total, but only about 1B do work on any given token. That ratio explains why generation is fast on modest hardware while the weights still need the memory of a 7B model.
The card says pre-training used 12.8 trillion tokens of public and synthetic data, run on the JUPITER, Jean Zay and MeluXina supercomputers. Post-training covered 36.9 million SFT samples and then four reinforcement learning stages, the final one aimed at agentic tasks of up to 75 turns. OVHcloud valued the public compute support at about €7 million in its announcement. The card estimates training energy at 1.16 GWh, or roughly 40 tonnes of CO₂e. Reasoning effort can be set to disabled, low, medium or high, with medium as the default.
Benchmarks and stated limits
OVHcloud's self-reported scores on the model card include 88.3 on HumanEval+, 75.1 on MBPP+, 50.5 on GPQA Diamond and 59.6 on BFCL V3, a function-calling benchmark. The model averages 68.0 across 12 benchmarks. The card names instruction following as Olala's strongest area and competition maths as its weakest. Nobody has published independent evaluations yet, and the press coverage on Next notes that OVHcloud's comparison against other open models appears only as an image.
The card lists limitations plainly. Olala holds less factual knowledge than larger models and can state errors with confidence. It works best on short, single-purpose requests. Its comprehension drops outside English and the major European languages, and long reasoning chains can fall into repetition loops. Context is capped at 32K tokens shared between input and output, and OVHcloud says a 128K version is planned for the end of 2026. On safety testing, the attack success rate fell from 12.5% at the SFT checkpoint to 6.3% in the final model. Over-refusal of benign prompts rose from 7.0% to 10.7%, and 15.0% of responses hit the token limit while still reasoning.
Running it locally: check the fine print
The hardware claims in the press release need reading alongside the docs. OVHcloud's announcement says the model runs on an 8GB GPU at 160 tokens per second on an RTX A4500, and at 50 tokens per second on a CPU alone, with no test conditions given. Next pointed out that the BF16 weights alone need about 14GB, so those figures imply quantisation. The GGUF repository lists a 13.6GB BF16 file, plus Q8, Q6_K and Q5_K_M quantisations between 6.0GB and 6.9GB. OVHcloud withdrew its 4-bit build because quality dropped measurably. The model card's own CPU figure is about 6 tokens per second on a 64-core server using plain Transformers, which is far slower than the headline number.
The bigger catch for local builders is tooling. Upstream llama.cpp, Ollama, LM Studio, official vLLM and mlx-lm can't load the olala architecture yet. The GGUF files need OVHcloud's llama.cpp fork, and GPU serving needs OVHcloud's vLLM fork or its ovhaillm/olala-vllm Docker image, which the card says requires an NVIDIA GPU with at least 24GB. The repository warns that a generic Q8 quantisation "can produce NaN on this model", so homemade quantisations are a risk.
On Apple Silicon the numbers look good. The GGUF page reports 72 tokens per second for Q6_K on a 16GB MacBook Air M5. Once the fork is built, the documented command for an OpenAI-compatible server on a 16GB Mac is:
./build/bin/llama-server -m Olala-v1.0-mini-7A1B-q6_k.gguf -ngl 99 -fa 1 -c 16384 --port 8080The -c value is shared across parallel slots and can't exceed 32,768. Owners of M1 machines should set OLALA_NO_FUSED_MOE=1, which the page says lifts Q5_K_M generation on an M1 Pro from 34 to 54 tokens per second.
What builders in New Zealand could do with it
The hosted options on OVHcloud AI Endpoints and Hugging Face Inference Providers run in the EU, according to the card. For NZ teams handling personal information, the Apache 2.0 weights matter more than the hosted service, because they let a team keep prompts on its own Mac or on a GPU server in a local data centre. A 1B-active model with tool calling and adjustable reasoning suits narrow jobs such as routing support tickets, extracting fields from forms, or acting as a cheap subagent inside a larger pipeline. OVHcloud's own reference customer uses Olala this way: Qwant runs the model for short "flash answer" responses and reports answers 20% faster at about a third of the previous running cost, per the press release. Those figures come from OVHcloud and haven't been independently checked.
Two cautions apply locally. The model card doesn't mention te reo Māori or Pacific languages, so teams that need either should test before relying on the model. Depending on a vendor fork also means waiting for upstream llama.cpp and vLLM support before Olala fits into existing Ollama or vLLM deployments. Until that support lands, we'd treat Olala as a model to benchmark against Qwen-class small models on your own tasks, and not yet as a drop-in replacement. Members who run those comparisons can share their results through ASAI projects.
Sources
- Olala-v1.0-mini-7A1B model card (Hugging Face)
- OVH announces Olala mini 7A1B (OVHcloud newsroom)
- Olala-v1.0-mini-7A1B-GGUF (Hugging Face)
- OVHcloud publie son premier modèle à poids ouverts, Olala (Next)
- OVHcloud launches first open-source AI model family (IT Brief)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.