Thomson Reuters says it spent 40 million dollars post-training an open-source base model, and releases a small version as open weights

Sources: primary, discussion
Summary: Thomson Reuters states that it spent 40 million dollars post-training an open-source base model into a model it calls Thomson, and that a small version is released as open weights on Hugging Face. The announcement names no base model, states the model was trained on less than 10 percent of Thomson Reuters content, and gives the first deployment as Tabular Analysis in CoCounsel Legal. The release carries capability claims from the CEO and two named academics but publishes no evaluation numbers or method, and the technical report it points at covers the underlying foundation model rather than Thomson, so this entry records the announced facts and not the capability claim.
Why it matters: The stated figure covers talent and compute, and is a public data point on what domain specialisation from an open-source base costs against frontier pretraining.