# 汤森路透推出首个大语言模型 Thomson，基于阿里 Qwen3.5-397B 构建

- 来源：X.PIN (@thexpin)
- 发布时间：2026-08-25 14:50
- AIHOT 分数：54
- AIHOT 链接：https://aihot.virxact.com/items/cmt8b7hrj3002ro7368rec1uy
- 原文链接：https://x.com/thexpin/status/2092142430728695939

## AI 摘要

汤森路透推出其首个大语言模型 Thomson，基于阿里 Qwen3.5-397B 构建，总研发投入约 4000 万美元。该模型与帝国理工学院联合重训以保障安全与政治中立，并用自有内容进一步训练。Thomson 在斯坦福 LegalBench 得分 0.823，在 Harvey 法律智能体基准上接近 Opus 4.8。

## 正文

Thomson Reuters has launched Thomson, its first LLM — built on Alibaba's Qwen3.5-397B. Total R&D: ~$40M (the $450K figure circulating online covers only the final training run).

Retrained with Imperial College London for safety, ethics, and political neutrality, then further trained on Thomson Reuters' own content.

Billed as a top-tier model: 0.823 on Stanford LegalBench (behind Gemini 3.1 Pro & GPT-5.5), nearly matching Opus 4.8 on Harvey's Legal Agent Benchmark. Caveat: Thomson used inference-time scaling; GPT-5.5 didn't. Deep search: 0.53 vs GPT-5.4's 0.65.

Why skip OpenAI/Anthropic? High long-term costs and limited customizability — a smaller in-house model fits document review better.