Computing Life · Share · YageAug 28, 2026, 08:00 UTC+8

The Third Path for Domain Models: A 175-Year-Old Company's $40M Answer

域模型的第三条路：一个 175 年公司的 $40M 答案

Thomson Reuters spent $40M to post-train Alibaba's Qwen open-weight model on legal data, claiming domain scores that beat Anthropic Haiku 4.5. Compute cost was only $100-200K; the real spend went into 175 years of proprietary case law, hundreds of expert annotators, and a high-resolution eval system. Three years ago Bloomberg burned far more cash training a 50B model from scratch that never shipped. Harvey later used full-parameter RL on GLM 5.3 Flash and reported beating GPT-5.5 and Opus 4.8 Max on legal benchmarks. The post flags two caveats: domain injection caused measurable regression on math and coding, and Chinese open-source vendors are tightening commercial licenses, so license review must now precede any base-model decision.

Why it matters: Thomson Reuters spent $40M building a legal-domain model on an open-weight base, self-reporting scores above Haiku 4.5, with open weights and cost transparency. This isn't a PR piece—it's a route analysis with concrete numbers and a Bloomberg failure comparison. Not scored hig...

Read the original ↗Export Markdown

Take

The reason to click is the cost breakdown: $40M total over two years, but only $100-200K on compute. The real spend went into 175 years of proprietary case law, hundreds of expert annotators, and a high-resolution eval system that can actually separate good answers from great ones. They picked Alibaba's Qwen3.6-35B-A3B as the base, injected less than 10% of their proprietary content via continued pre-training, and reported a legal domain score of 75.2, beating Haiku 4.5 at 67.7.

I'd discount the scores a bit since they're self-reported and the test set isn't fully public. But the architecture logic matters more than the numbers: open-weight base handles general reasoning, RAG carries dynamic facts, and the proprietary data flywheel focuses purely on evaluation and behavioral alignment. Three years ago Bloomberg burned far more cash training a 50B model from scratch that never shipped. Harvey took a similar path this June, using full-parameter RL on GLM 5.3 Flash and reporting wins over GPT-5.5 and Opus 4.8 Max on legal benchmarks.

The post flags two caveats worth noting: domain injection caused measurable regression on math and coding, and Chinese open-source vendors are tightening commercial licenses, so license review has to happen before any base-model decision. That licensing risk might be more dangerous than any model performance gap.

Source: Computing Life · Share · Yage · yage.ai
Published: Aug 28, 2026, 08:00 UTC+8 · Aug 28, 2026