Thomson Reuters invested 40 million dollars in its own frontier-class model
The media giant with a market capitalization of over 80 billion dollars and 24 thousand employees around the world refused cloud solutions and created its own artificial intelligence system.
The company recruited a team of 30 specialists, used 368 GPU B200 and released the Thomson-1.0 model in three months. The foundation consisted of the open checkpoints Qwen3.5-397B (MoE with 17 billion active parameters) and Qwen3.6-35B.
The models were named Thomson-1.0-Large and Thomson-1.0-Small. The cost of fine-tuning the Large version was less than 450 thousand dollars, and the entire project cost 40 million dollars, a significant part of which went to infrastructure and research.
According to aggregated tests, Thomson Large showed 78.5%, surpassing Claude Sonnet 5 (75.7%). The leadership was held by Claude Opus 4.8 with 79.5%.
How the "model factory" system worked
Development included three sequential stages. At the first stage, the model was retrained under corporate values: first, undesirable behavior was removed by direct weight correction, then it was fixed through LoRA adaptation.
The second stage is intermediate training on 200 billion tokens from the company's closed corpus, which amounted to less than 2% of the original 19 trillion. The remaining data were discarded.
The final stage included training a deep research agent with its own set of organizational tools using DPO and RL. Separate training with a teacher was not conducted to not lose the general abilities of the model.
An important point — after the intermediate training, the weights were merged with the original checkpoint. This technique proved to be more effective than regularization through LoRA, reduced learning rate or short schedules.
Comparison with the Imago approach: scaling principles
Thomson Reuters applied the same principle "model as part of the organization". Their system repeated three layers: values, knowledge and skills. LoRA was used only at the initial stage of adaptation, then the weight merging and final polishing followed.
The company consistently selected data: proprietary documents were used, along with their synthetic retellings in the form of lessons and replays against forgetting. The principle "data is more important than volume" worked at maximum power.
In contrast to coding, where the skills of writing code were slightly impaired, the model significantly improved work with long context and agent tasks. This confirms that the approach scales from small businesses to large corporations.
What this means for large companies
Thomson Reuters has proven that creating its own frontier-class model has become a real option at reasonable costs for a corporation. Instead of constant renting of cloud solutions, companies can fully own their system.
For large organizations with their own code, data and the ability to collect a dataset, the full weights after the initial adaptation become a profitable alternative. The main thing is to start with the organization's dataset: it will survive any base model.