Thomson 1.0: As Foundation Models Become Infrastructure, Enterprises Need to Own the Capability to Produce Professional Intelligence

 In August 2026, Thomson Reuters released the Thomson 1.0 model family. The publicly available Thomson-1.0-Small was continually trained from Qwen3.6-35B-A3B, with approximately 35 billion total parameters and around 3 billion parameters activated per inference step. The production-grade Large model, meanwhile, was derived from a larger Qwen-family foundation model.

The overall program involved approximately $40 million in cumulative investment, covering research and engineering talent, compute, domain experts, data, and supporting engineering infrastructure. Yet the GPU cost of the final full training run for the Large model was less than $450,000.

On Thomson Reuters’ internally constructed aggregate legal-domain evaluation, the Small model scored 75.2, compared with 72.7 for its underlying Qwen model. A closer examination, however, shows that Thomson 1.0 does not lead across every conventional legal benchmark. Its more substantial improvements appear in tasks that more closely resemble actual legal work, including Human Queries, Deep Research, document processing, RAG, and agentic workflows.

To interpret this simply as “a traditional company trained its own legal LLM” would therefore underestimate its significance.

The more important lesson for enterprise AI transformation is a broader shift: as general-purpose foundation models become stronger and increasingly open, enterprises no longer need to recreate general intelligence from scratch. They can instead concentrate capital on proprietary data, expert judgment, evaluation systems, workflows, and continual-learning capabilities—and gradually build their own capacity to produce professional intelligence.

From Buying Model Capability to Building Intelligence Production Capability

The enormous difference between the approximately $40 million overall investment and the few hundred thousand dollars required for a final training run helps explain the economics of Thomson 1.0.

The expensive part is not the last GPU run. It is building a system capable of repeatedly producing high-quality professional models: organizing domain data, defining tasks, coordinating expert annotation, designing evaluations, generating preference data, training models, identifying failures, retraining, and safely deploying improved versions into production.

Thomson Reuters reportedly filtered roughly 200 billion tokens from more than 19 trillion candidate tokens, then further shaped model behavior through continual pre-training, preference optimization, reinforcement learning, domain ontologies, Deep Research environments, and expert-generated questions.

The most important lesson for enterprises is therefore not the training technique itself. It is the changing definition of a data asset.

Owning large volumes of PDFs, contracts, policies, operating procedures, and knowledge bases does not automatically create an AI moat. The real value lies in transforming those materials—and the work performed by professionals around them—into:

professional questions, correct answers, failure cases, preference comparisons, evaluation criteria, task trajectories, tool-call records, and human corrections.

Raw data improves machine capability only when it is converted into trainable and evaluable supervision signals.

Enterprise AI Is Shifting from “Knowledge Injection” to “Work Capability Formation”

The evaluation results of Thomson 1.0 reveal another important lesson.

Its most significant improvements do not appear exclusively on traditional legal examinations. They are more visible in tasks resembling actual professional work. This suggests that the objective of domain-model optimization should not simply be to make a model “know more professional knowledge.”

The more important questions are whether the model can understand a task in the way a professional would, search for and use relevant sources correctly, structure outputs according to professional standards, cite supporting evidence, and ultimately produce a work product that can be passed into the next stage of a business process.

This changes how enterprises should evaluate AI systems.

An enterprise Agent should not be judged only by accuracy or benchmark scores. Evaluation should also include task completion rate, human-edit rate, rework rate, professional acceptance rate, supervision time, completion cycle time, and the cost per qualified work product.

The ultimate optimization target for enterprise AI is therefore work performance, not model scores.

Owning a Model Does Not Mean Abandoning Claude, GPT, or Other Frontier Models

Another sign of maturity in Thomson Reuters’ strategy is that owning Thomson 1.0 has not led the company toward a single-model architecture.

CoCounsel remains multi-model by design. Thomson is initially deployed in areas such as Tabular Analysis, where workloads are frequent and structured and where Thomson Reuters has strong proprietary data advantages. More complex long-tail tasks continue to rely on leading external models.

This provides enterprises with a more practical technology strategy.

Tasks that are high-frequency, stable, data-rich, cost-sensitive, and governed by clear evaluation criteria can progressively migrate toward enterprise-owned models. Frontier reasoning, complex long-tail problems, and rapidly evolving capabilities can continue to be sourced from external model providers.

Enterprises therefore do not need to choose between “build” and “buy.”

What they need is an intelligence infrastructure capable of dynamically selecting models according to quality, cost, latency, security, specialization, and task complexity.

Continual Learning Is Not About Letting AI Rewrite Itself

Continual Learning is another concept surrounding Thomson 1.0 that can easily be overstated.

It does not mean that every time a lawyer edits an answer, the model immediately updates its own parameters. Reliable enterprise continual learning should operate as a controlled loop:

production systems generate data; failures are identified; experts review cases and convert them into training data; models are retrained offline; regression evaluations, safety checks, and compliance controls are performed; and validated models are then introduced into production through staged rollout.

What enterprises should automate, therefore, is not “models evolving by themselves,” but the process of:

Discover Problems → Capture Data → Create Supervision → Validate Improvement → Deploy Safely.

The organizations that can make this loop faster, less expensive, and more reliable will be the ones that genuinely develop continual-improvement capabilities.

The Real Decision-Making Lesson from Thomson 1.0

For the vast majority of enterprises, the Thomson case does not mean that now is the time to start training a proprietary large language model.

A more rational progression remains:

Context → RAG → Tools → Agent Workflow → Eval → Failure Data → Post-training.

Model training becomes economically justified only when an enterprise has demonstrated that its primary performance bottlenecks originate in the model’s intrinsic behavior, that these failures are frequent, stable, and measurable, and that the organization possesses sufficient proprietary data, expert resources, and workload scale.

The most important industry implication of Thomson 1.0 is therefore not that “domain models are back.”

It shows that as foundation models increasingly become standardized infrastructure, competition in enterprise AI is moving one layer deeper:

General intelligence can increasingly be bought. Professional intelligence must be accumulated.

The most valuable long-term investment is therefore not any single model, but an Enterprise Intelligence Factory capable of continuously converting enterprise data, knowledge, rules, expert judgment, and real-world work feedback into machine capability.

Models will change. Foundation models will be replaced. GPU costs will continue to fall.

But once an enterprise establishes a closed loop of:

Data → Supervision → Evaluation → Model/Agent → Workflow → Human Feedback → Re-learning

it no longer merely owns an AI product.

It owns a professional intelligence production capability that can continue to accumulate value across generations of models.

Related Topic