$450K TO TRAIN A FRONTIER AI MODEL?
Thomson Reuters CTO Joel Hron explains why the company built Thomson, a domain-specific frontier model adapted from open-weight checkpoints like Qwen 3.5, using decades of proprietary legal and tax content and expert judgment. The conversation covers the training lifecycle from realignment and continual pre-training to agentic RL, the economics ($450K final run versus roughly $40M total), team structure, CoCounsel orchestration, verification challenges in legal and tax, and the future balance between open-weight and frontier models.
Joel Hron, CTO of Thomson Reuters, joins the conversation to explain why the company built its own frontier model, Thomson, on top of open-weight checkpoints rather than relying solely on commercial APIs. He begins by framing the strategy around what he calls the AI flywheel: compute, data, and expertise. As frontier gains increasingly depend on expert judgment and out-of-distribution data rather than generic internet text, Thomson Reuters argues its proprietary content and its roughly 3,000 domain experts in law, tax, and compliance are its real differentiators.
Hron describes the journey as more than two years long. The company started with basic LoRA-style fine-tuning through commercial provider APIs, but found catastrophic forgetting made models overly narrow: performance on the targeted evals improved while general capabilities degraded. That pushed the team toward a full lifecycle methodology spanning mid-training, continual pre-training, post-training, and reinforcement learning. An acquisition (Safe Sign) seeded the team, which is roughly 30 people in science and engineering. Hron stresses that the bet on open-weight models paid off: the open community kept pace within a few months, so Thomson Reuters could build on top of checkpoints instead of pre-training from scratch.
The initial checkpoint was Qwen 3.5, a model of about 397 billion parameters, chosen because it was the best performer on their benchmarks at the time of the training run in late May or early June. Hron notes that the pipeline is designed to be adaptable, so newer checkpoints such as GLM 5.3 Flash can be swapped in. The first step was model realignment. Thomson Reuters wanted to take any open-weight model, whose provenance can be hard to validate, and realign its weights to values the company can stand behind, grounded in the roughly 150-year-old principles of journalistic truthfulness and objectivity that apply equally to law and financial services.
On economics, Hron separates the roughly $450,000 final training run compute cost from the approximately $40 million total spent over two years across people and compute. He argues this is modest relative to what enterprises spend on token consumption annually, and that the goal is a pipeline cheap enough to re-run in weeks rather than months on future checkpoints.
The data work is described as the majority of the effort. Content is primarily textual, drawn from legal archives, editorial material, and regulatory tax content, with structured and numerical accounting data a future direction. Thomson Reuters has used less than 10% of its raw content for continual pre-training, selected through evaluation-driven decisions rather than random sampling, and worked with Datology on this phase. Experts enrich the corpus with metadata, headnotes, and citation linkages built over decades. Fine-tuning and RL data come from preference judgments by in-house experts, and the company has decomposed its content products such as Westlaw and Practical Law into agentic tools that the model learns to call natively. Hron compares this to frontier labs training models to use web search or file system tools.
The post-training stack combines DPO-style preference optimization driven by expert judgment with an agentic RL loop in which the model operates inside a harness with access to these tools. Infrastructure is built with Lambda Labs for training, and the company partners with AWS, GCP, and Microsoft for hosting and inference. Hron declines to detail the orchestration stack beyond the technical report on Hugging Face.
The product context is CoCounsel, an agentic system built on Claude’s Agent SDK that also draws on OpenAI, Gemini, and Thomson models. Thomson currently handles roughly 5 to 10 percent of inference, mainly for high-volume bulk document review such as M&A due diligence. Hron doubts the share will ever reach 100 percent, since some frontier capabilities should remain with frontier labs, but expects it to grow. SLMs are already deployed in the stack, and Thomson could serve as a distillation source. Hron also sees value in customers consuming the model directly through an API.
The conversation then turns to tax and verification. Hron notes that Thomson Reuters’ tax engine is effectively the verification tool for tax calculations, though tax work also involves substantial judgment about how expenses should be treated. The Ready to Review product uses AI in a loop to extract data from documents, compute a 1040 return, and verify outputs iteratively. On legal work, the discussion covers what CoCounsel can do, from research memos and brief drafts to contract negotiation against playbooks and M&A risk review, and the difficulty of verifying citations that are real but mischaracterized. Hron argues that the human lawyer’s role shifts toward higher-order strategy, while the risks of automation bias are too high to tolerate. He also suggests junior lawyers may get more repetitions across more deals, even as grunt work disappears.
On continual learning, Hron pushes back on the common meaning of the term as online weight updates from production signals. For Thomson Reuters it means adding proprietary knowledge to an open model while managing catastrophic forgetting, with customer data kept private.
The closing exchange addresses the broader thesis. The host argues open-weight models will absorb most routine work while frontier labs keep complex problems. Hron agrees with the framing but adds that harness development, serving, and distributed computing frameworks are part of the open ecosystem, and that enterprise optionality and sovereignty matter. The discussion draws a parallel to Linux and open databases overtaking proprietary incumbents during the dot-com era.