Thomson Reuters自研LLM評測：45萬美元練出法律AI模型 | Thomson Reuters LLM Review: $450K Buys a Legal AI

By Kit 小克 | AI Tool Observer | 2026-08-27

🇹🇼 Thomson Reuters自研LLM評測：45萬美元練出法律AI模型

近日法律科技圈最熱門的話題，是Thomson Reuters正式推出自家研發的大型語言模型「Thomson 1.0」——這是傳統法律／金融資訊巨頭首次跳下場自建法律AI模型，而不是單純向OpenAI或Anthropic租算力貼牌。這篇評測用no-hype角度，拆解這次企業自建LLM熱潮背後的真相與限制。

45萬美元練出一個法律LLM？

Thomson Reuters這兩年砸了約4000萬美元在人才與算力上，但最終這一版Thomson 1.0的訓練成本只要45萬美元。關鍵在於它不是從零開始訓練，而是拿開源模型（最新底層是阿里巴巴的Qwen 3.5）持續微調，過程中甚至換了六次底層開源模型以追上最新技術。訓練資料來自Westlaw、Practical Law、Checkpoint與路透社新聞庫，還加入數百位內部法律專家的知識，但官方坦承目前只用了公司內部素材的不到10%。

真的贏過GPT-5.4和Claude Sonnet 5嗎？

Thomson Reuters在自家「Deep Research」測試中比較Thomson、GPT-5.4與Claude Sonnet 5：純網路搜尋情境下，Thomson表現不算最強；但只要加入TR自家法律資料庫，Thomson就「贏過」另外兩家。問題是，這份基準測試由TR自己設計、自己執行，目前沒有任何第三方獨立驗證的benchmark報告——公司只承諾「這週內」會公布更詳細的技術數據。換句話說，「贏過前沿模型」這句話目前只能算是廠商自己的一面之詞。

CoCounsel還是多模型架構，不是全面換血

Thomson 1.0目前只會先用在CoCounsel Legal的Tabular Analysis（表格化文件審閱）功能上，其他任務CoCounsel仍然會呼叫GPT、Claude等第三方前沿模型。也就是說，這不是「用自家AI取代OpenAI」的宣示，而是「哪個模型在哪個任務上比較划算，就用哪個」的務實混合策略。

對其他企業的啟示

如果你手上有夠深、夠獨特的垂直領域資料，微調開源模型的成本可能遠比想像中低
但「贏過前沿模型」的宣稱在有獨立評測之前都該打折扣
目前釋出的開源權重版本很小，且僅限非商業學術用途，開發者平台也還很早期，一般開發者短期內用不到

Thomson Reuters這次示範了一條「不用燒幾億美元也能做出可用垂直LLM」的路，但這條路的終點是不是真的比租用GPT-5.4或Claude Sonnet 5划算，還需要真正的第三方數字說話。詳見Thomson Reuters官方新聞稿與LawSites的詳細報導。

好不好用，試了才知道。

🇺🇸 Thomson Reuters LLM Review: $450K Buys a Legal AI

Thomson Reuters just did something most legal-tech vendors talk about but rarely ship: it built and shipped its own large language model, called Thomson 1.0, instead of just reselling GPT or Claude with a legal skin on top. Here is the honest, no-hype rundown of what it actually is — and what its beats-the-frontier-models claim really means.

A $450K Training Run, After $40M in R&D

Thomson Reuters spent roughly $40 million over two years on people and compute to get here, but the final training run for the shipping version of Thomson 1.0 reportedly cost about $450,000. The trick: it did not train from scratch. Thomson is a continually fine-tuned version of an open-weight base model — most recently Alibaba Qwen 3.5 — and the team swapped the underlying open model about six times as better options shipped. Training data comes from Westlaw, Practical Law, Checkpoint, and Reuters news, plus input from hundreds of in-house legal experts. TR admits it has used less than 10% of its available proprietary content so far.

Does It Actually Beat GPT-5.4 and Claude Sonnet 5?

In TRs own Deep Research benchmark, Thomson trailed GPT-5.4 and Claude Sonnet 5 on web-only tasks but outscored both once TRs proprietary legal content was added to the mix. The catch: this benchmark was designed and run entirely in-house, with no independent evaluation published yet — TR says a fuller technical report is coming this week. So for now, the beats-the-frontier-models claim is a vendors own homework, graded by the vendor.

Still a Multi-Model Product, Not a Replacement

Thomson 1.0 is launching narrowly — inside CoCounsel Legals Tabular Analysis feature for structured document review. CoCounsel stays multi-model: third-party frontier LLMs still handle everything Thomson is not specifically better at. This is not we replaced OpenAI, it is we route each task to whichever model is cheapest and good enough.

What This Means If You are Building on Vertical Data

If you sit on deep, differentiated domain data, fine-tuning an open-weight model may cost far less than you would assume
Discount any beats-the-frontier-models claim until independent benchmarks exist
The open-weight release is small and non-commercial/academic-only for now, and the developer portal is early — not something most teams can build on yet

Thomson Reuters just proved a cheaper path to a usable vertical LLM exists. Whether that path actually beats renting GPT-5.4 or Claude Sonnet 5 long-term is still an open question until real third-party numbers show up. See Thomson Reuters official announcement and Artificial Lawyers analysis for more.

好不好用，試了才知道。

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士