Thomson-1.0-Small from Thomson Reuters: 3 billion active parameters, 35 you still have to keep in memory

On 18 August 2026, at 15:33 UTC, the thomsonreuters organisation published Thomson-1.0-Small on Hugging Face: not an academic experiment, but an open-weight model built on the group’s proprietary content — decades of news, contracts, regulatory filings, case law — for legal, tax and journalism work. For anyone assessing a sector-specialised model to run inside their own perimeter, this is the kind of release to check line by line before it enters a tender: how much it actually weighs, what is needed to run it, what the card declares — or leaves out — about the data it learned from. We verify it here, as of 26 August.

The facts, in order

Publication: repository created on 18 August 2026 at 15:33:05 UTC, a minor update on 24 August. Organisation thomsonreuters, verification badge active (isVerified: true) — the only other model under the same account is an academic release from 2023, all but unnoticed: this is, in effect, the group’s first weights release.
Access: gated: false. Verified without authenticating: requesting a weight shard and the licence file both return a direct redirect to the content, no login required.
Adoption, today: 214 downloads, 109 likes — the numbers of a model published eight days ago, not of a release already filtered by the market.
Weight: 35.1 billion total parameters, 3 billion active per token (mixture-of-experts architecture, Qwen3.6 family), BF16 weights across sixteen safetensors files, 70.2 GB — 65.4 GiB — per the repository’s public metadata.
Base: built on tri-fair-lab/Snowdon1.1-Small (Apache 2.0), itself derived from Alibaba’s Qwen/Qwen3.6-35B-A3B (Apache 2.0). The chain is traceable back to its origin; the licence changes only at the last step.
Licence: PolyForm Strict 1.0.0 — non-commercial use, no redistribution or derivative works without a separate agreement with Thomson Reuters; the base checkpoint was permissive, this one is not.
Documentation: the technical report is not on arXiv and has no persistent identifier. It is a PDF hosted in a Hugging Face Space belonging to tri-fair-lab — unverified (isVerified: false) — the joint lab with Imperial College London that also signs the base checkpoint. Link checked today, working.
Context: 262,144 tokens declared “natively” — an exact 256K (256×1,024), not the “262K” you would round to on a quick reading of the figure.

What it is, in practice

Thomson-1.0-Small does not start from scratch: it is an open-weight Qwen checkpoint, realigned on values through a document the card calls the Public AI Constitution — declaredly public and modifiable, not a proprietary value system — then put through continual pre-training on Thomson Reuters content. The card mainly claims efficiency in training — 1.63×10²³ FLOP, 35,207 B200 GPU-hours — not in production serving, on which it says nothing.

The benchmark tables — read, not reproduced — show the model ahead of its own base checkpoint on most legal, tax and deep-research domain entries, with a narrower margin on maths and coding. The comparison is against models chosen by the vendor — Qwen3.6-35B-A3B, Gemma 4-31B, Claude Haiku 4.5 — measured, the table states, on the same harness: a vendor’s own figures, not an independent trial.

The sum the README does not do

The card declares no minimum hardware configuration: no recommended GPU, no memory threshold. The only certain figure is the one above: 70.2 GB of weights on disk. Working out the memory needed to serve it takes a step the card does not take.

A mixture-of-experts model chooses, for every token, which experts to activate: here three billion parameters out of thirty-five, the number that decides compute cost per token and the one the card highlights. It is not, however, the number that decides the memory needed to run it in production: different tokens in the same batch of requests route to different experts, so serving the model without stalling on the wrong expert requires keeping all thirty-five billion parameters ready in fast memory, not just the three billion active ones. This is a known property of Mixture-of-Experts architectures, not a claim made by this card. Add to that figure, once requests approach the declared context of 262,144 tokens, the attention cache — another chapter of cost no line of the document estimates.

Downloadable weights without the compute that makes them runnable are not sovereignty: here the missing sum is not the three billion the card invites you to reason with, but the thirty-two more left out of the count — the same gap already measured on a model of comparable size, a sign this is not an isolated case.

Training data: what is there, what is missing

On this point the card is more forthcoming than the sector average — worth comparing with the near-total silence we usually measure. It states that continual pre-training “ingests Thomson Reuters’ proprietary data — decades of news, contracts, regulatory filings, case law, statutes and practitioner guidance”, and specifies the composition: “The mid-training corpus comprises 200B tokens curated from a pool of over 19T tokens of permissively public and proprietary data”, split roughly evenly between proprietary documents, their synthetic rephrasings, and general-capability replay data.

That is more than almost any competitor declares. Still missing: a date range for the proprietary corpus, a word on personal data inside the case law cited, and, above all, the corpus itself — of the three dataset repositories published by partner lab tri-fair-lab, none is training material, they are the evaluation sets cited in the tables. The material Thomson-1.0-Small actually learned from stays, in full, inside the perimeter of whoever wrote it.

The family the name implies

The card introduces itself this way: “It is the open-weight member of the Thomson-1.0 family.” The name implies other members, not open. None is documented: no link, no stated size, no note on what extra capability, if any, sets them apart from these downloadable weights. On Hugging Face the organisation shows only one other repository, the 2023 academic model, unrelated to this family. If a larger or managed version exists, it cannot today be verified from a public source.

What we do not know

We have not tried the model on our own documents. The scores in the tables are computed by the vendor, with a harness and comparison models chosen by the vendor: they are not an independent trial. We offer no legal qualification of the PolyForm Strict licence, nor of what, in practice, falls within the non-commercial use it permits. The memory figure needed to run the model is our own reasoning from the parameter count, not a claim made by the card. And among the repository’s files sits a video_preprocessor_config.json the card explains nowhere — the document discusses text only, never images or video: a minor inconsistency, but one to ask about before installing, not to ignore.

How we solve it

For a model like this, verification does not stop at the licence: it enters the model register as a dated line — version, file digest, channel and date of acquisition, the text of the PolyForm Strict licence archived exactly as it read that day, the approved use for legal or tax work and who in the organisation approved it — ready to produce at an inspection, a tender, a board meeting. The same system holds models, data, contracts, archives and documents together in a single operational model, on which AI agents execute decisions with an operator in command: for large enterprises, defence, public administration and healthcare. On-premise on autonomous machines that require no deep integration into the client’s network, or on a dedicated cloud with a dedicated VPN and a data centre in Italy — always with shared management: you do not need to already have, in-house, someone who administers these systems.

See the service · Talk to an engineer

Do you need to assess a model specialised for legal or tax work, with nobody in the organisation able to say how much memory it actually takes to run? Half an hour with one of our specialists, at no cost: we bring the real estimate of hardware, licence and declared data before it goes into a tender.