DatologyAI Launches Curation Studio, Claims 6x Less Compute

DatologyAI launched Datology Curation Studio on 9 October 2026, a product that runs the company’s data-curation pipeline for teams training their own language models. In a companion post the company says that, in its own experiments, models trained on curated data matched the largest baseline model’s score with 6× less training compute.

The launch post on X described it as “customizable frontier data curation for every team training its own models” and drew about 3.7 million views by 11 October. The figures below are DatologyAI’s, from experiments it ran on public datasets.

What DatologyAI Launched

The company says Curation Studio delivers “our whole curation engine as a straightforward, guided experience, with presets and automation of our algorithms”, and that more advanced teams can fork its recipes to run their own experiments.

It runs inside the customer’s cloud. The launch post says Curation Studio “is deployed inside your VPC”, with the curation core, data, lineage and curated outputs staying in the customer’s account.

The first release covers LLMs across web, math, code and multilingual data. The post lists four stages: clean, calibrate, create (synthetic data through rephrasing) and compose (mixing the subsets into a training set).

The Customers It Names

The launch post names Thomson Reuters, HRT, Deepgram, Arcee and Unconventional AI as production customers, alongside “several of the Mag 7”, without naming those. The results post names Thomson Reuters, Hudson River Trading and Deepgram.

The company gives one customer example: it says it helped Thomson Reuters curate data to mid- and post-train a Qwen model into Thomson-1, which it says outperformed frontier models such as GPT-5.6 Sol in blind comparisons on legal workflows, “for 100x lower cost per task.”

The Experiment Behind the 6x Claim

DatologyAI says it built a baseline by hand-mixing 39 publicly curated datasets, including Nemotron releases, The Stack v3, FineWeb and DCLM, following published recipes from OLMo 3, Nemotron 3 and IBM Granite 4.1. It then ran the same sources through Curation Studio with default settings and trained models at five sizes with identical recipes.

At the largest size, the company reports that a post-trained 30B-A3B model on curated data scored 46.8% against 37.7% for the baseline on mean benchmark score, with both pretrained for 1 trillion tokens.

It reports AIME 2025 rising from 19.4 to 35.6, HumanEval++ from 54.1 to 70.2, and MATH-500 from 76.4 to 89.6. It also says a 12B-A1.4B model trained on 440B tokens of curated data outscored the 30B-A3B baseline on overall math and code with roughly a fifth of the pretraining compute.

In pretraining, the company reports a 10-12% improvement in average bits-per-byte across five model sizes spanning 3.9e19 to 3.7e22 in pretraining compute.

How It Says the Pipeline Works

The company says it has processed more than 23PiB over three years, most of it in 2026. For its most recent experimental curation it says it processed 355TiB through about 670TiB of intermediaries to produce a final 26TiB.

On math data, the results post says the product found repetition leakage in widely used sources such as MegaMath and UltraData-Math, with accidental repetitions “at times exceeding 1,000×”.

The Structural Read

The company frames the choice for model builders as whether to “Keep renting somebody else’s intelligence, or own it yourselves.” Curation Studio is pitched at the teams that choose to own, starting from data they already hold.

The comparison DatologyAI reports holds the sources and the training recipes fixed and changes only the data processing. The company says its baseline drew on 39 publicly curated datasets, including Nemotron releases.

As we read it, deploying inside the customer’s VPC is part of the pitch as much as the algorithms: the launch post says the curation core, data, lineage and curated outputs stay in the customer’s account.

Three Implications

TEAMS TRAINING THEIR OWN MODELS DatologyAI says the first release covers LLMs across web, math, code and multilingual data, with presets for teams that want curation to just work.

BUYERS COMPARING DATA VENDORS The company’s figures come from its own experiments on public data; the posts describe the setup but do not include independent replication.

ANYONE PRICING IT Neither post, nor the product page as read on 11 October, states a price; the company points to a demo request and a live session on 27 October.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The essay describes “The Corpus Moat” as “built by pretraining, eroding as the public web is exhausted and synthetic data closes gaps”, and “The Container Moat” as “built by closed data loops inside a customer’s environment, nascent but the deepest position in the stack.”

As we read it, Curation Studio sits between the two: it reworks public corpora with synthetic rephrasing, and it is deployed inside the customer’s own cloud to run on the data that customer holds.

What Is Not Established

Neither post, nor the product page as read on 11 October, states a price. Access is through a demo request, and the company says it will hold a live session on 27 October.

The 6×, the 46.8% and the benchmark gains come from DatologyAI’s own experiments on public data; the posts describe the setup but do not include independent replication. The Thomson-1 comparison is the company’s account of a customer project.

The posts say automated mixing will come “later this year” and that new data domains and modalities are planned; they give no dates beyond that.

The Bottom Line

DatologyAI has turned its data-curation pipeline into a product that runs in the customer’s own cloud, and it backs the launch with company-run results: a 46.8% against 37.7% mean score at the 30B-A3B scale, and a claim that curated data matched its largest baseline with 6× less training compute. No price is published, and the results have not been independently reproduced.