Thomson Reuters Spent $40M on a Model. The Final Run Cost $450K.

TL;DR

On August 24, Thomson Reuters launched Thomson, its own legal and tax frontier model. The company puts the total spend at about $40 million over two years across people and compute, and the cost of the final training run at roughly $450,000. It is not trained from scratch: it is a continual-learning stack built on Alibaba's Qwen open weights, value re-aligned with Imperial College London, then fed decades of Westlaw and Reuters archives. The open-weight version is a 35B mixture-of-experts model that cost 35,207 B200 GPU-hours end to end. And after all that, CoCounsel Legal still routes most work to Anthropic and OpenAI.

Where the $40 million went, and where it didn't

The two numbers together are the whole story. Forty million dollars is a real corporate AI program: researchers, data engineers, hundreds of subject-matter experts writing evaluation criteria, two years of iteration. Four hundred and fifty thousand dollars is the compute bill for the run that produced the shipping artifact.

Per the model card for Thomson-1.0-Small, the full pipeline consumed roughly 1.63 x 10^23 FLOP over 35,207 B200 GPU-hours. At public cloud rates for a B200, that is a number a well-funded Series B could cover. Thomson Reuters did the work in partnership with Lambda for compute and DatologyAI for data curation.

The ratio is the lesson. In frontier training the run is the cheap part; the expensive part is knowing what to put in it. That is not a new observation, but it is rare to see an incumbent publish both halves of the bill.

The base is Chinese, so there was a step for that

The open-weight Thomson-1.0-Small is, by its own card, obtained by "repurposing" Qwen3.6-35B-A3B. Reporting from The Decoder and LawSites says the larger production model sits on Qwen3.5-397B-A17B, and that Thomson Reuters swapped its starting checkpoint roughly half a dozen times during the project.

Before any legal content went in, the team ran a value re-alignment pass with Imperial College London, producing an intermediate checkpoint named Snowdon, after the mountain in Wales. The target is not a house style guide; it is the Public AI Constitution, an openly published document, applied through Constitutional DPO and again as a conformance reward during reinforcement learning.

The card reports the effect on its political-neutrality eval: the stock Qwen3.6-35B-A3B scores 78.5, Snowdon 91.5, and Thomson 98.5. Think of it as buying a used car with the previous owner's radio presets still saved: the engine is fine, but you change the stations before you hand the keys to a client.

The benchmark it wins, and the famous one it loses

Specialization shows up exactly where you would expect and nowhere else. On the Harvey Legal Agent Benchmark, the tool-using agentic test, Thomson-1.0-Small posts 73.4 against 69.5 for the raw Qwen base and 60.5 for Claude Haiku 4.5. Document processing and RAG is the widest gap in the card: 78.8 versus 43.8 for Haiku.

Then there is Stanford LegalBench, the benchmark every legal AI deck cites. Thomson-1.0-Small scores 79.9. Its own pre-legal-training ancestor Snowdon scores 80.9. The untouched Qwen base scores 80.3. Google's Gemma 4-31B, which has never seen a Westlaw headnote, wins the row at 83.1.

That is not a failure, it is a statement about what these benchmarks measure. LegalBench is mostly short classification and issue-spotting tasks a general model already handles. The value of domain training shows up in long-document work with citations to a corpus you own, which is exactly what the Harvey, Deep Research, and RAG rows capture. The card also shows the tax gap is bigger than the legal one: 78.6 versus 68.0 for the base on tax Deep Research.

They built their own model and kept paying for everyone else's

Thomson goes live first inside Tabular Analysis, the high-volume document-review feature in CoCounsel Legal, where it is the default and switchable in admin settings. That is one feature. CoCounsel stays a multi-model product: Thomson where domain training wins, third-party frontier models everywhere else.

This is the honest version of the build-versus-rent argument, and it deserves more credit than the usual sovereignty press release. CTO Joel Hron framed it to reporters as renting a house versus buying one. The nuance the framing hides is that they bought a house and kept the hotel reservation for most nights.

The open weights are not open

Here is the part builders should read twice. Thomson-1.0-Small ships under PolyForm Strict 1.0.0, which grants the right to run the model for noncommercial purposes and nothing else. No distribution. No modification. No derivative works. You cannot fine-tune it.

Meanwhile Snowdon1.1-Small, the intermediate checkpoint it came from, is Apache 2.0, as is the Qwen base underneath. The lineage gets less permissive as it gets closer to the money, which is at least internally consistent.

So the honest read of "you can do this too" is: you can read the recipe, you can start from the same free base, you cannot start from theirs.

What the recipe actually is

The technical report, Continual Learning of Frontier Models for Sovereign AI, is public, and the pipeline is reproducible in outline:

• Value re-alignment first, via Constitutional DPO against a published constitution, before any domain data goes in.
• Continual pre-training on 200B tokens curated from a pool of over 19T, split roughly evenly between real proprietary documents, synthetic rephrasings of those documents, and general-capability replay data.
• Model merging during CPT to protect general skills, which is why the general-benchmark table barely moves.
• Post-training on preference data derived from expert-authored content, including ontology-driven pairs built from schemas like IRAC for case law. The claim there is sharp: preference data mined from material your editors already wrote is data no annotation budget can buy.
• Agentic RL inside the actual product tools, with rewards for faithful tool use and correct citations.

The general-capability table backs the anti-forgetting claim: MMLU-Pro 85.7 versus 85.2 for the base, GPQA-Diamond 85.4 versus 85.2, SWE-bench Pro 34.4 versus 34.3. Two regressions are visible, AIME 2026 at 90.0 down from Snowdon's 93.3 and Terminal-Bench 2.1 at 40.5 against the base's 45.2, so the forgetting is reduced rather than eliminated. Those evals were run through the UK AI Security Institute's Inspect framework, which is a better sign than a homemade harness.

Caveats

• Every benchmark here is Thomson Reuters running its own harness. The general-capability rows use Inspect, but the legal and tax evals are in-house, and the Deep Research aggregate is a weighted blend of completeness, factuality, relevance, and coherence chosen by the vendor.
• The published benchmark table covers the 35B open-weight Small, not the larger production model actually powering CoCounsel. Numbers for the flagship come from press briefings, not a card you can check.
• $450,000 is the final training run. It is not the cost of the six abandoned base swaps, the expert time, or the two years of salary that produced the data pipeline.
• Robustness is the soft row: Thomson lands 56.3, up from the base's 48.7 but well behind Haiku 4.5 at 70.2. Domain depth did not buy resilience.

Key Takeaways

• $40 million over two years, $450,000 for the final run. The compute for a domain frontier model is now the smallest line item; the data and the experts are the moat.
• 35,207 B200 GPU-hours and 1.63 x 10^23 FLOP produced a 35B/3B-active MoE with a 262,144-token context, built by repurposing Qwen rather than pre-training.
• Specialization is narrow and measurable. Thomson beats its own base by 3.9 points on the Harvey agent benchmark and loses to it on Stanford LegalBench. Pick the benchmark that matches your actual workload.
• A value re-alignment pass on a Chinese base is now a shipping step, not a talking point: political neutrality moves 78.5 to 91.5 before any legal data is added.
• "Open weight" is doing heavy lifting. PolyForm Strict 1.0.0 forbids modification and redistribution, so the release is downloadable, not forkable.
• Owning a model does not end the API bill. Thomson powers one feature by default; CoCounsel keeps routing everything else to third-party frontier models.

Sources: Thomson Reuters press release, Hugging Face, thomsonreuters/Thomson-1.0-Small model card, Hugging Face, tri-fair-lab/Snowdon1.1-Small, Thomson 1.0 Technical Report, LawSites, The Decoder, SiliconANGLE, Thomson Reuters, next-generation CoCounsel Legal, PolyForm Strict License 1.0.0