Legal News · Topic
Model Releases legal news
Proprietary legal AI models, and what they were trained on.
- Model Releases
Thomson Reuters launches Thomson, its own legal LLM
Thomson Reuters now runs a model it owns, named Thomson, instead of relying only on outside frontier vendors. It starts from an open-source base and is specialised on the company's own libraries, Westlaw and Practical Law among them, plus Checkpoint and Reuters material, which the company says amounts to under a tenth of its content so far. It reaches customers first through the Tabular Analysis feature in a coming CoCounsel Legal release, so the buying questions are whether vendor-owned inference lowers the price of high-volume review work and whether the accuracy claims hold up outside benchmarks the vendor ran itself.
- Model Releases
Harvey previews Tenet, its first post-trained legal model
Harvey has taken an existing open-weight base, Kimi K3, and post-trained it with Fireworks into a legal-specific model it calls Tenet. On its own Legal Agent Benchmark, Harvey reports the model finishes close to double the held-out tasks of the untuned base and 20 percent more on the contracts set, and it claims about a tenth of the previous cost per cell in table review. This is a research preview with no ship date, no pricing and no published weights, so nothing changes for buyers today, but it signals that the largest legal AI vendor is reducing its dependence on frontier labs, which bears on what firms eventually pay for high-volume document work.