17 November 2026 · 4 min read
OVHcloud Launches Olala Mini 7A1B: An Open MoE Model That Runs on 8 GB GPU
OVHcloud announced its first open-source LLM, Olala mini 7A1B, a mixture-of-experts model with 7B parameters that runs on 8 GB of GPU memory. No benchmarks yet, but early deployment shows speed and cost gains.
What was announced
On 17 November 2026, European cloud provider OVHcloud unveiled Olala mini 7A1B, the first model in its new Olala family of large language models [1][2]. The model is a mixture of experts (MoE) with 7 billion total parameters, of which roughly 1 billion are active per request [1][3]. It will be released under the Apache 2.0 licence on Hugging Face, with weights, a pre-trained base model, a technical report and AI Act compliance documentation [2][4]. The official launch presentation is scheduled for the OVHcloud AI Summit on 19 November [2].
What is new
Olala mini is designed to run on modest hardware. OVHcloud states it needs as little as 8 GB of GPU memory and can also run on a CPU without a GPU [1][4]. The company reports throughput of 160 tokens per second on an NVIDIA RTX A4500 (a 20 GB workstation card) and 50 tokens per second on an unspecified processor [1][3]. The context window is 32,000 tokens, with a planned extension to 128,000 tokens in version 1.1 by the end of 2026 [1][2].
The architecture is unusual. As an MoE, the model has 256 experts per layer, and a router activates only a subset for each input, which OVHcloud says reduces memory use and speeds up inference [1][3]. The attention mechanism is hybrid, combining Differential-TPA (a transformer attention variant) with Mamba-3 (a state-space model) [1][2]. This mix is intended to handle long sequences efficiently while keeping the model compact.
Training was carried out entirely on three European public supercomputers—Jupiter (Germany), MeluXina (Luxembourg) and Jean Zay (France)—using GH200, A100 and H100 GPUs [1][2]. OVHcloud puts the public support at the equivalent of €7 million in compute hours [1][3]. The model’s development stems from the acquisition of Dragon LLM (formerly Lingua Custodia), a Paris-based AI lab, in March 2026 [3].
Early deployment and claimed efficiency
French search engine Qwant, which is part of the same corporate group as OVHcloud, has already adopted Olala mini for its “flash answer” feature [1][3]. OVHcloud says Qwant saw answers delivered 20% faster and operating costs cut by a factor of three compared with an unnamed previous model [1][2]. This provides a real-world example of the cost and speed benefits, though the comparison is not independent.
What it means for developers
For developers and builders, Olala mini offers a few concrete possibilities:
- Local and private deployment. With an 8 GB GPU requirement, the model can run on a single gaming or workstation card, or even on a CPU. This makes it feasible to keep inference in-house, avoiding data leaving your infrastructure—useful for regulated industries or sensitive applications [1][4].
- Agentic applications. OVHcloud explicitly positions the model for agentic tasks, where a model uses tools and completes multi-step workflows [1]. The low latency (160 tokens/sec) could suit interactive agents.
- Open-source flexibility. The Apache 2.0 licence and full release of weights and base model allow fine-tuning, inspection and integration into custom pipelines without licensing restrictions [2][4].
- European sovereignty. The model was trained on European public infrastructure, which may matter for organisations with compliance or geopolitical concerns [2][4].
However, several points remain unverified. No benchmark scores have been published, so the model’s quality relative to other small open models (such as Ministral, Gemma or Qwen) is unknown [1][3]. The 8 GB memory claim is puzzling: 7 billion parameters in BF16 format occupy about 14 GB. OVHcloud has not explained how it fits into 8 GB—likely through quantisation or partial loading of experts, but this is not detailed [3]. The speed figures were measured on an RTX A4500 with 20 GB, not on an 8 GB card, so real-world performance on the minimum spec is unclear [3]. Developers should treat these numbers as provisional until independent testing is possible.
Roadmap
OVHcloud plans to release version 1.1 by the end of 2026, extending the context to 128,000 tokens and improving mathematical reasoning [1][2]. A much larger model, version 2.0 with over 100 billion parameters, is on the roadmap, with specialisations in enterprise AI agents, cloud DevOps and cybersecurity [1][2]. These future models are expected to power OVHcloud’s own products, such as the encrypted OVHai Workspace, which will require models that run locally on user machines [1][3].
Bottom line
Olala mini is a pragmatic, hardware-efficient open model that lowers the barrier to running LLMs locally. Its true capabilities will only become clear once the weights are released and the community evaluates it against established benchmarks. For now, it represents a deliberate step by a European cloud provider to build its own AI stack, with an emphasis on cost control, data sovereignty and open access.
Sources
- OVHcloud launches Olala mini, an open model that runs on 8 GB of GPU memory
- OVHcloud Introduces Olala Mini 7A1B as Its First Open-Source LLM
- OVHcloud unveils Olala mini, its first LLM: speeds, no scores
- OVHcloud launches first open-source AI model family
- ovhcloud
- olala
- open-source
- llm
- mixture-of-experts
- local-ai
Researched and drafted with AI from the sources above.
Learn to build with AI
Hands-on tutorials on Cursor, Claude Code, AI agents and more.