OVHcloud announces Olala mini 7A1B ahead of open release | Magica
OVHcloud unveils Olala mini, with public compute support and benchmarks pending
Editorial Team
••📖7 min read
The French cloud provider plans to release Olala mini 7A1B on November 19, backed by acquired expertise and European public computing support. Its first deployment offers cost claims, but no published benchmarks yet establish its competitive quality.
OVHcloud will release its first language model on Hugging Face under Apache 2.0, with an official launch scheduled for November 19.
The model builds on an acquired AI specialist and public computing support valued by OVHcloud at €7 million; the project’s total cost is undisclosed.
Claimed savings at Qwant come from a business linked to OVHcloud’s founder; independent quality comparisons remain absent.
OVHcloud, the French provider of website hosting and public and private cloud services, announced its first family of large language models on November 17, 2026. The first, Olala mini 7A1B, is intended to run on modest hardware and eventually underpin the company’s own software. It gives the infrastructure supplier a model it controls, although its announcement does not yet establish how well that model answers questions.
The company says it will make the weights and pre-trained base model available under the Apache 2.0 licence, alongside a technical report and AI Act compliance documentation. The public release on Hugging Face, a platform for distributing AI models, is expected on November 19, when OVHcloud holds its AI Summit. The announcement therefore precedes the evidence developers need to assess it.
OVHcloud's company-reported specifications show seven billion total parameters, including one billion active during processing. Source: OVHcloud.
A cloud supplier buys the expertise to own a model
Octave Klaba, who founded OVHcloud in Roubaix in 1999, is its chairman and chief executive. His argument for building models is about dependence as well as sales: the company cannot assume other laboratories will keep releasing their models openly. In its announcement, OVHcloud says it already serves 1.7 million customers across 46 data centres, giving the project an existing route to business users.
This strategy is being pursued by a company still controlled by its founding family. Its shareholder disclosure puts the Klaba family’s holding at 74.27% of share capital as of October 23, 2026.
New Street Research has initiated Anthropic at Buy ahead of its planned IPO. Its $105 target bets on expanding AI demand, while the filing shows surging revenue and improved customer-serving economics alongside volatile operating results and vast infrastructure commitments.
Editorial Team
Olala also has a history before OVHcloud. Its architecture comes from Dragon LLM, formerly Lingua Custodia, a Paris company founded in 2011 around financial translation. OVHcloud acquired it in March and renamed it OVHai LLM, according to a report on the launch. A review of its history traces its roots and its June 2024 win in the European Commission’s Large AI Grand Challenge. The project therefore draws on an existing language-technology team, rather than starting with infrastructure alone.
OVHcloud says in its announcement that Olala mini was trained from scratch on Jupiter, MeluXina and Jean Zay, public supercomputers in Germany, Luxembourg and France. It values the support at €7 million in computing hours, without identifying the allocation scheme. That is a valuation of public resources, not a disclosed cash training bill or the project’s total cost.
The company brings an operating business to this effort. Its FY2026 results show €1,145.1 million in revenue and €9.4 million in net income for the year ended August 31, 2026. Those group figures establish its scale and profitability; they do not show whether building models is profitable.
In an interview accompanying the launch, Klaba said the company did not intend to raise funds for the strategy and had fewer than 100 people working on AI across three Paris laboratories. His claim that training costs have fallen twentyfold over the past two to three years comes without a common workload or methodology to test the comparison.
OVHcloud's company-reported generation speeds measure throughput, not answer quality. The RTX A4500 result does not establish speed on the minimum 8 GB configuration. Source: OVHcloud.
Local execution has a commercial purpose
According to OVHcloud’s specifications, Olala mini has seven billion parameters, with one billion active during processing, distributed across 256 experts per layer. This mixture-of-experts design uses selected parts of the model rather than all its parameters for each step. OVHcloud combines Differential-TPA attention with Mamba-3 in a hybrid architecture intended to reduce inference compute and memory demands.
The company claims operation with as little as 8 GB of GPU memory, generation at 160 tokens per second on an RTX A4500, and 50 tokens per second on an unspecified standard processor without a GPU. These figures describe generation throughput, not answer accuracy or successful completion of business tasks.
The minimum-memory claim also needs qualification. As the technical analysis points out, the A4500 has 20 GB of memory. Seven billion parameters in the announced BF16 format, at two bytes each, imply about 14 GB for weights alone. The release needs to explain whether the 8 GB configuration uses compression, partial loading or another arrangement; the headline speed cannot simply be assigned to that minimum configuration.
Local execution matters to a specific product plan. OVHcloud intends Olala to power OVHai Workspace, its suite of messaging, storage and videoconferencing tools presented in June. Its planned encrypted version will require models running directly on users’ workstations, according to the company’s announcement and the account of the workspace strategy.
Dragon LLM was one of three acquisitions this year, alongside encryption specialist Seald and voice-AI developer Gladia. In the launch interview, Klaba also outlined programming assistants for businesses, comparable to Claude Code, Codex and Mistral Vibe. The company targets 100 million users without specifying a deadline. These are product ambitions, rather than demonstrated adoption of Olala.
Compact, locally deployable models are already a competitive category. Mistral AI, the French model developer, introduced Ministral 3B and 8B in October 2024 for on-device and edge computing, including agentic workflows. The licensing comparison needs care: that 2024 release offered self-deployment through commercial licences, with Ministral 8B Instruct weights available for research use. Olala’s promised Apache 2.0 release would permit reuse and modification, including commercially, as the launch specifications explain.
Small size and local execution alone therefore do not establish an advantage. Google’s Gemma and Alibaba’s Qwen are also alternatives identified in the launch coverage. No published Olala benchmarks yet support a quality ranking, and OVHcloud’s announcement gives no hosted-service price for a comparable cost assessment.
Qwant offers a use case with an ownership caveat
OVHcloud says in its deployment account that the French search engine Qwant already uses Olala for its flash-answer feature, which generates summaries for search results. It claims answers are 20% faster and operating costs are one-third of those under the previous model.
The previous model is unnamed, and the announcement does not supply workload, traffic, hardware or answer-quality controls. The figures therefore cannot establish a general price-performance advantage.
There is also a relationship behind the reference. Qwant belongs to Synfonium, owned 75% by Octave and Miroslaw Klaba and 25% by Caisse des Dépôts, according to the ownership account. The deployment is evidence of use, but it is not an arm’s-length endorsement by an unrelated customer.
OVHcloud's company-reported roadmap proposes increasing context capacity from 32,000 to 128,000 tokens; the larger window remains planned. Source: OVHcloud.
What developers need from the November 19 release
OVHcloud’s next steps are already broader. Version 1.1, expected by year-end, is intended to expand the context window—the text a model can consider in one interaction—from 32,000 to 128,000 tokens and improve mathematical reasoning. A later version 2.0 is planned to exceed 100 billion parameters, with specialisations in enterprise agents, cloud development and operations, and defensive and offensive cybersecurity, as the roadmap describes. No release date is given for that larger version.
For customers considering the first model, the November 19 weights and technical report need to establish which configuration fits in 8 GB, document the throughput tests and disclose how the model was trained. The decisive comparison is operating cost at a consistent workload and answer quality. Until that evidence is available, OVHcloud has demonstrated a strategy for owning its models, while the commercial advantage of this first release remains unproven.
The Alphabet-backed drug developer has announced $7.1 billion in funding in six months. Its new capital will expand proprietary AI models and clinical infrastructure as company-reported preclinical progress awaits testing in patients.
Microsoft enters Ignite with a previously announced Copilot redesign that combines office applications and autonomous agents. Availability, usage-based charges and reliability will determine what that expansion means for customers, while October's missing paid-seat update leaves adoption harder to assess.
A proposed Equity–Pact agreement would give UK film and TV performers consent and payment rights over digital replicas. Members must approve the package, which leaves explicit protections against AI training unresolved.
US stock futures fell before Nvidia's first earnings report since Jensen Huang's departure, with cash generation, tighter export controls and rate-hike expectations clouding the AI trade ahead of Anthropic's expected market debut.
Nvidia has created a standing board committee to oversee national security and export compliance after co-founder Jensen Huang’s departure. Investigation findings and operational changes remain unresolved.
Microsoft's CEO says team sizes, workflows and skills will change as AI reshapes engineering. About 60,000 employees are eligible for voluntary exits, but uptake and the prospect of later layoffs remain unresolved.
Hong Kong officials say financial firms and their human overseers remain responsible for AI agents. With 36 projects selected for sandbox testing, the unresolved task is turning that principle into reliable identity checks, delegated authority and human intervention.
Meta says stronger AI checks and India-specific tuning keep more suggestive and violent content out of Instagram recommendations. Its reporting metric leaves exposure unmeasured, while a government request to change ranking systems drew objections over the ministry’s authority.
OpenAI says NATO has used up to 300 billion raw input and output tokens a month for cyber defence. Access through NATO’s technology agency and an AI-agent exercise shows uptake, while costs, comparative performance and NATO-specific security results remain undisclosed.
Numa’s Series C finances a move from customer conversations into sales and service operations. Automaker distribution gives it reach, but vendor approvals and company-reported savings leave the broader operating-system pitch to prove.
Google’s research tool can turn uploaded sources into interactive dashboards, study aids and trackers. Shared links require full notebook access, personal inputs stay separate, and generation draws on compute-based usage limits.
Records detail Chinese data-center investments linked to Hexi founder Kang Wei, husband of former Megaspeed chief Alice Huang. Filings document financing and Huawei business, but do not establish a route for restricted Nvidia chips.
ElevenLabs plans to acquire Kotoba, the startup behind Japanese speech technology used in Salesforce’s Agentforce Voice. The deal adds a 15-person team and deployed products to its existing Japanese operation, but financial terms and integration plans remain undisclosed.
Tesla plans €3 billion of expansion at its profitable Grünheide factory to build Cybercabs from 2029, eventually targeting up to one million annually. The driverless taxi still lacks German and EU approval, and its capacity target differs from an older plan for the whole site.
Our World in Data’s comparison finds negative views of AI exceed positive ones across five probability-based surveys. The primary results also reveal substantial mixed views, acceptance of some applications and a sharper divide over who controls development and bears its local costs.
NVIDIA's Nemotron for Science pairs an existing open model with nearly 50,000 chemistry questions and scientific post-training recipes. Partner results suggest scope for specialized agents, but the reported accuracy and inference savings do not establish the full cost of adapting them.
AMD says the Lux AI supercomputer is open to scientific users at Oak Ridge. Half its annual capacity supports the Genesis Mission and half is reserved for paid access, including AMD development, while workload limits and undisclosed rates shape its practical value.
Apple is adding app-by-app and folder-specific permissions, repeat approval requests and clearer warnings in macOS 27.3 beta 2. Existing permissions remain intact, limiting the immediate reach of the privacy changes.
OpenAI reportedly discussed buying Goodfire for more than $5 billion, over four times its February financing valuation. No deal resulted. A disclosed paid assessment agreement offers cooperation without ownership, while the technology’s safety benefits remain to be demonstrated.