Thomson Reuters Thomson-1: Run the Open Legal AI Model Locally
Thomson Reuters open-weighted a 35B legal AI model built on Cohere. Hands-on tests show it catching contract red flags and citing real tax law.
What is Thomson-1 and why does it matter?
Thomson-1 is an open weight legal AI model built by Thomson Reuters, the company behind Westlaw and much of the legal research infrastructure lawyers already rely on. It’s based on Cohere’s 35 billion parameter mixture of experts model, retrained on decades of Thomson Reuters’ proprietary legal, tax, and news data. Unlike a general purpose chatbot with a legal prompt bolted on top, it’s purpose built for tasks like contract review, legal research, and tax analysis, and because it’s released with open weights, anyone with enough GPU hardware can download it and run it themselves instead of going through an API.
TL;DR
- Thomson-1 is a 35B mixture of experts model built on Cohere’s open weight architecture, retrained by Thomson Reuters on its own legal, tax, and news archives.
- The model is fully open weight, meaning it can be downloaded and run locally with tools like vLLM rather than accessed only through a hosted API.
- In a hands-on contract review test, the model caught all seven deliberately planted red flags in a mutual NDA and flagged an eighth issue the tester hadn’t planted.
- Thomson Reuters designed the model around query sufficiency, meaning it’s trained to ask clarifying questions instead of guessing when a legal query lacks enough context.
- The training pipeline includes a dedicated deep research reinforcement learning stage, aimed at reducing hallucinated citations, which showed up in a tax law test where the model self-corrected a factual error mid-answer.
- Running the small variant locally consumed roughly 87GB of VRAM in the tester’s configuration, though that figure depends heavily on KV cache settings and won’t match every setup.
- The model is aimed at high-stakes professional use, where knowing the limits of what it can answer matters as much as producing a fluent response.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How was Thomson-1 built?
According to the model’s own documentation as covered in the hands-on test, Thomson-1 was built through a three-stage pipeline on top of Cohere’s open weight base model:
Value focus. Before any legal knowledge gets added, the open weight base model goes through a realignment stage using constitutional DPO (direct preference optimization). This step shapes the model’s behavior and values ahead of domain training.
Knowledge focus. Continued pre-training injects Thomson Reuters’ proprietary legal, tax, and news data into the model. That version is then merged back with the original base model, a step meant to add domain expertise without degrading the model’s general reasoning ability.
Behavior and agentic focus. The final stage runs two more rounds of DPO plus two rounds of reinforcement learning. One of those reinforcement learning rounds is described as “agentic end-to-end deep research,” trained using an actual research harness rather than just prompt engineering. This is the stage Thomson Reuters points to as the source of the model’s citation accuracy and reduced hallucination rate.
The result is a model that isn’t simply a general LLM told to “act like a lawyer” in its system prompt. The legal behavior is trained into the weights across multiple stages, which is also why the model can be evaluated on tasks that require judgment, not just recall.
What happened in the contract review test?
To test whether the model could function as a first-pass reviewer, a mutual NDA was written with seven deliberate red flags embedded in the text, including a one-sided indemnity clause that left all the risk with one party, and a global non-compete clause hidden inside otherwise ordinary language. These are the kinds of clauses that are easy to miss on a quick read and expensive to discover later.
The model was given the full contract with a single instruction: review it like a lawyer would and flag the riskiest clauses for the receiving party. It caught all seven planted issues and identified an eighth problem buried in the obligations clause that hadn’t been intentionally inserted. The model closed its output with a summary, structured the way a lawyer’s internal review memo might read, after working through the document clause by clause.
Does the model know when it doesn’t have enough information?
This is where the test moved from knowledge to judgment. The model was given a single vague prompt: “Can I get out of this contract?” No contract was attached, no jurisdiction specified, and no context given.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
A model trained mainly to sound helpful will typically produce generic, made-up-sounding advice in that situation. Thomson-1 instead asked six clarifying questions, refused to guess at a jurisdiction or contract type, and explicitly warned against unilaterally stopping performance on a contract without a valid legal basis. Thomson Reuters refers to this behavior as “query sufficiency,” and it appears to be a deliberate, benchmarked property of the model rather than an accident of prompting. The reasoning behind the response ran to roughly 3,300 tokens and took about 21 seconds to generate in the test environment, reflecting a model that spends more compute thinking through a problem before answering, similar to how a lawyer would pause before giving advice on an incomplete fact pattern.
Does Thomson-1 hallucinate on legal and tax facts?
To test citation accuracy, the model was asked a simple, checkable factual question: current US capital gains tax rates, with a request to cite sources. This tests the raw training data rather than any live research tool, since the setup used direct model inference rather than a retrieval pipeline.
The model correctly flagged its own knowledge cutoff instead of guessing at current-year numbers, cited real Internal Revenue Code sections, and reasoned through the implications of the Tax Cuts and Jobs Act (TCJA) sunset provisions rather than inventing a specific future rate. Partway through its answer, it caught its own error on the pre-TCJA top tax rate and corrected itself in place. That kind of self-correction is consistent with the model’s dedicated deep research reinforcement learning stage, which was built specifically to reduce hallucinated citations in legal and tax contexts.
How do you run Thomson-1 locally?
Because Thomson-1 is released as an open weight model, it can be run outside of any hosted API using inference frameworks built for large mixture of experts models. The hands-on test used vLLM, a popular open source inference engine, to serve the model locally and query it through a simple script that hit the local endpoint.
Running the small variant of Thomson-1 consumed approximately 87GB of VRAM in that test configuration, though the tester noted this was partly a function of a high KV cache allocation on the GPU system used, meaning actual VRAM needs will vary based on context length, batch size, and inference settings. That VRAM footprint puts the model well beyond consumer GPU territory and squarely into workstation or multi-GPU server territory, whether that’s owned hardware or a rented GPU instance.
Is Thomson-1 worth it for legal AI work?
For teams that already fine-tune open models for legal domain work, Thomson-1 changes the calculus. Rather than starting from a general purpose base model and layering on legal fine-tuning, teams can start from a model that has already been through value alignment, legal and tax domain training, and agentic research reinforcement learning. Further fine-tuning may still make sense for specific jurisdictions or client-specific requirements, but the baseline capability appears strong enough that generic legal fine-tuning may no longer be necessary for many use cases.
The bigger signal is that a major legal information company chose to open-weight a model this capable rather than keeping it locked behind an API. That gives developers, legal tech startups, and in-house teams a way to build on top of a domain-specific model without recurring API costs, provided they have the infrastructure to host it.
Frequently Asked Questions
What company built Thomson-1?
Thomson Reuters, the company that owns Westlaw and a large share of commercial legal research infrastructure, built Thomson-1 on top of Cohere’s open weight 35 billion parameter mixture of experts model.
Is Thomson-1 free to use?
One coffee. One working app.
You bring the idea. Remy manages the project.
Thomson-1 is released with open weights, which means it can be downloaded and run on your own hardware or rented GPU infrastructure without per-query API fees, though you still need to cover the compute cost of hosting it.
How much VRAM does Thomson-1 need to run?
In one hands-on test running the small variant through vLLM, VRAM usage reached approximately 87GB, driven in part by a high KV cache setting. Actual requirements will vary depending on quantization, context length, and inference configuration.
What is “query sufficiency” in Thomson-1?
Query sufficiency is a behavior Thomson Reuters specifically trained and benchmarked into the model: instead of guessing an answer to a vague or underspecified legal question, the model asks targeted clarifying questions before offering advice.
Can Thomson-1 replace a lawyer for contract review?
The model performed well on a controlled contract review test, catching all planted red flags plus an extra issue, but it’s designed and marketed as a tool for first-pass review and legal research support, not as a replacement for licensed legal judgment on real matters.