OVHcloud Opens a 7-Billion-Parameter AI Model That Runs on an 8GB Graphics Card, as Cloudflare Frees Its Clairic Extraction Models and RadixArk Trains a 27B Agent on Just 1,200 Tasks, Showing Small Open-Weight Machine Learning Is Getting Cheap Enough to Own

Three days of releases have made one point hard to miss: the most practical machine learning news this week is not about the biggest models, but about the smallest ones. On Thursday, November 19, European cloud provider OVHcloud used its AI Summit to launch Olala mini 7A1B, its first in-house language model, with open weights that can run on an 8GB graphics card or even a regular processor. A day earlier, Cloudflare published open weights for two data-extraction models, Clairic and Clairic-flash, and the startup RadixArk released DaisyRose-27B, an agent model it post-trained on only about 1,200 carefully chosen tasks. SenseTime added SenseNova Loop-9B, a 9-billion-parameter model it says competes with systems three times its size.

Together, these launches show how fast compact, downloadable models are improving, and why that matters for any company weighing what to rent from a frontier AI lab and what it can run on its own hardware.

OVHcloud’s Olala mini: a 7B model built to run almost anywhere

OVHcloud first announced Olala on November 17 and set the official launch for its November 19 AI Summit. According to the company, Olala mini 7A1B holds about 7 billion parameters but only around 1 billion are active for each request, spread across 256 experts per layer and combining Mamba-3 layers with attention layers. That design keeps memory needs low. OVHcloud says the model runs on a GPU with as little as 8GB of memory, reaching 160 tokens per second on an RTX A4500 card, and about 50 tokens per second on a standard CPU. The context window is 32,000 tokens for now, and the weights are released under the permissive Apache 2.0 license.

The model was trained on three European public supercomputers, JUPITER, MeluXina and Jean Zay, with compute time the company valued at about €7 million. A detailed review by ChainAIWire notes that the training run covered about 12.8 trillion tokens and that every benchmark figure so far comes from OVHcloud itself, including a 68.0 average across 12 tests and 88.3 on the HumanEval+ coding benchmark. Independent checks have not yet caught up.

The most useful data point for businesses is a customer result. French search engine Qwant already uses Olala mini for its quick-answer feature, and OVHcloud says answers arrive 20% faster while running costs fell to a third of the previous model’s. CEO Octave Klaba argued in the announcement that training costs have dropped about twentyfold in three years, making it possible to build models profitably instead of relying on other labs to keep sharing theirs. A version 1.1 with a 128,000-token context window is promised by year-end, and a version 2.0 above 100 billion parameters is planned after that.

Cloudflare opens Clairic for documents, receipts and forecasts

Cloudflare took a narrower route. In a November 18 blog post, it said its updated 27-billion-parameter Clairic and a new 9-billion-parameter Clairic-flash are now available as open weights on Hugging Face and as hosted endpoints on Workers AI. These are not chatbots. They take text, images or PDFs plus a list of fields and return typed answers such as a supplier name, line items, totals, estimates or forecasts, along with confidence scores.

Cloudflare says a redesigned architecture improved accuracy, cutting image-reading errors on its internal benchmark from 207 to zero, and the models can now read up to four images or PDF pages per request. Pricing on Workers AI is $0.24 per million input tokens for Clairic and $0.10 for Clairic-flash, with no charge for output tokens. Until December 1, 2026, both cost $0.10. Unite.AI reports that the two models are built on Qwen base models and ship with the code needed to run them locally.

RadixArk shows quality data can beat brute-force compute

The third release speaks to how these models are made. RadixArk says it built DaisyRose-27B on top of Qwen3.8-27B using reinforcement learning on long, multi-step agent tasks. Out of thousands of candidate tasks built with partner Eigent AI, about 1,200 made the final cut, each designed to target a specific mistake the base model kept making. Training ran for 18 days on 144 Nvidia H200 GPUs, roughly 63,000 GPU-hours in total.

The result more than doubled the model’s pass rate on the difficult Terminal-Bench 4.1 test, from 5.6% to 12.0%. That is still a low absolute score, which is a reminder that long, autonomous coding work remains hard for models of this size. RadixArk also says its training backend needed 57% less GPU time than a comparable open-source setup, and its sandbox infrastructure cost about 80% less than popular cloud sandbox providers. All checkpoints, including compressed FP8 and NVFP4 versions, are free to download.

SenseTime squeezes more thinking out of fewer parameters

SenseTime’s official release of SenseNova Loop-9B takes yet another approach: it reuses the same middle layers twice, turning 32 physical layers into 56 layer computations. The company says its 4-bit version takes up only about 4.9GB and can run in under 10GB of GPU memory, while supporting a context window of up to 512,000 tokens. On SenseTime’s own tests, the 9B model scored 65.31 on the AIDABench data-analysis benchmark versus 65.01 for the much larger Qwen3.8-27B.

Why it matters for business

For most companies, the expensive part of AI is not the occasional hard question sent to a frontier model. It is the millions of routine calls behind the scenes: pulling numbers from invoices, sorting tickets, summarizing records, deciding which tool should handle a request. This week’s releases target exactly that layer. When a capable model fits on an 8GB card or a laptop, three things change.

  • Cost control. Qwant’s reported two-thirds cost cut and Cloudflare’s input-only pricing show how quickly per-task costs fall when a smaller model is good enough.
  • Data control. Open weights under Apache 2.0 let companies keep sensitive documents on their own servers, which matters for regulated industries and for European buyers working under the EU AI Act. OVHcloud released AI Act compliance paperwork alongside its model.
  • Less lock-in. Businesses can test, fine-tune and swap these models without being tied to one provider’s roadmap or price changes.

There are caveats. Nearly every benchmark cited this week comes from the companies selling the models. OVHcloud itself warns that a model with roughly 1 billion active parameters stores less factual knowledge and can state wrong facts with confidence, so it should be paired with tools and trusted sources. And ChainAIWire notes that popular local runtimes such as Ollama and LM Studio cannot yet load Olala without OVHcloud’s own forks.

The bottom line

The frontier labs will keep grabbing headlines with trillion-dollar valuations and ever-larger models. But the steady arrival of small, open, cheap-to-run models from cloud providers and startups is what will decide how widely machine learning spreads through everyday business software. The smart move for most firms now is to route routine work to compact models they control and save the premium models for the problems that truly need them.

Sources: OVHcloud, ChainAIWire, Cloudflare, Unite.AI, RadixArk, SenseTime.

Put AI to work on every call with MAJ.COM Voice AI

Smaller, faster AI models are making always-on automation affordable, and MAJ.COM Voice AI brings that to your phone lines: 24/7 AI voice agents that answer every inbound call and handle outbound sales and customer support, so your business never misses a lead. Get started with MAJ.COM Voice AI.

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous


Discover more from QUE.com

Subscribe to get the latest posts sent to your email.

Coron @QUE.COM

Editor at QUE.com Intelligence.

coronquecom has 196 posts and counting.See all posts by coronquecom

Leave a Reply

Connect with

Discover more from QUE.com

Subscribe now to keep reading and get access to the full archive.

Continue reading