European open-weight LLMs, sparse architectures and source availability

OVHcloud unveils Olala mini, a European 7B MoE with 1B active

The cloud company promises Apache-2.0 weights and a hybrid Mamba-3 design, but the public model repository and vendor speed claims still require independent checks.

Open Model Weights published19 Nov 2026
Primary sourceOVHcloud — official announcement of Olala mini 7A1B on November 17
Source published2026-11-17
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original minimalist editorial illustration representing model in AI research; conceptual, not actual model output.
AI-generated editorial illustration · Not official model imagery.

European cloud operator OVHcloud has announced its first in-house language-model family, Olala, with Olala mini 7A1B set for a formal presentation at its AI Summit on November 19. Unlike another generic API announcement, the project is explicitly framed around European training capability, a compact mixture-of-experts architecture and a promised open-weight release. The distinction between a public plan and a verified checkpoint remains important as the launch progresses. Primary source ↗

OVHcloud describes a model with around seven billion parameters in total but only one billion active per token. The design reportedly uses 256 experts per layer and hybrid attention combining Differential-TPA with Mamba-3. Its present context window is advertised at 32,000 tokens, with an expansion to 128,000 planned for version 1.1 before the end of 2026. The 7B total is still relevant for file size and loading: sparse activation is not proof that the model can be deployed at the memory footprint of a conventional 1B checkpoint.

The company's performance figures are ambitious. OVHcloud says Olala mini can run on an eight-gigabyte GPU and reaches 160 generated tokens per second on a named RTX A4500 system, while a CPU setup achieves 50 tokens per second. These are manufacturer statements without a complete independently checked workload, batch size, quantization path or public reproduction report. Open Model Weights has not tested those hardware combinations or established peak VRAM from the press release.

The European supply-chain angle is more concrete. OVHcloud names the Jupiter, MeluXina and Jean Zay public supercomputers as its training infrastructure and values their computing support at the equivalent of roughly seven million euros. The company also says search provider Qwant has adopted Olala for a flash-answer function, reporting 20% quicker answers and one-third of the previous running cost. Those operational gains are customer or publisher claims, not externally controlled evaluations.

OVHcloud promises Apache-2.0 licensing, weights, a pretrained base model, a technical report and AI Act compliance documentation on Hugging Face. Its new official fine-tuning guide points to an Olala repository identifier, but in this editorial review a version-specific public weight inventory, license file and pinned revision were not independently readable. We therefore treat availability as announced, not yet verified at the artifact level. A direct model profile should follow only after that evidence exists.

The significance of Olala mini is a European operator attempting to own part of the open-model training stack rather than relying solely on imported foundation weights. The next decisive test is practical: publication of inspectable files and reproducible deployment settings, followed by independent evaluation of quality, latency and licensing. Until then, the launch is a substantial commitment, not a confirmed local performance record.

OMW REGISTRY WATCH

What this changes for the evidence layer

As of the morning of 19 November 2026 the official announcement says weights will be available on Hugging Face under Apache-2.0 and the launch is scheduled for the Nov 19 AI Summit. A version-matched Hugging Face file manifest, license file and exact commit SHA were not independently accessible in this review. Do not label Olala as verified downloadable open weights or invent successful local benchmarks until those artifacts are checked.

PRIMARY SOURCE

OVHcloud — official announcement of Olala mini 7A1B on November 17

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗