2026年10月8日,LangChain 喺官方 blog 宣佈推出 LangSmith Fine-Tuning 同配套 CLI 工具 smithtune,將 LangSmith 入面保存嘅 agent trajectories(代理執行軌跡)一路變做微調數據集、訓練出定製模型,再回傳評測。呢次係 LangChain 喺「模型後訓練」(post-training)範疇嘅獨立產品線,唔係 Deep Agents 嘅例行版本更新。
究竟賣乜
smithtune 係一個 command line 工具,官方話佢「處理整個細分流程」:
- 由 LangSmith trajectories 直接產生同整理數據集
- 接 Fireworks 或 Baseten 做訓練(Baseten Loops 係 fully managed 嘅訓練基礎設施)
- 訓練完嘅模型由 LangSmith 做 evaluation,結果自動 upload 返 LangSmith 落分析
你可以直接打 CLI,亦可以喺 coding agent 入面裝 smithtune skill,令 coding agent 去驅動成個流程。官方話「Try LangSmith Fine-Tuning on GitHub」,skill 由 repository 安裝。
smithtune 現階段支援嘅係 supervised fine-tuning(SFT,監督式微調):畀模型 inputs/outputs 做範例,佢更新權重去模仿嗰行為。官方強調 LangSmith 嘅 trajectory 格式係「為後訓練設計」,記錄模型喺每一回合實際「見到」嘅 context,包括長程 agent 入面 deferred tool loading 導致嘅工具可用性變化——呢啲細節簡單 export 最終對話會全部丟失。
呢個嘢解決乜痛點
官方直接寫:「This gives teams a way to train specialized models without building the data pipeline by hand.」即係你唔使手搓數據管道,將生產環境跑過嘅 trace 變成訓練數據。
Fireworks 產品負責人 Pranav Jain 話:「One of the biggest drivers of fine-tuning gains is data selection. With the launch of smithtune, it's easier to connect that loop — from curated production traces in LangSmith to managed training and a served model on Fireworks. Teams can move seamlessly from data to training to deployment without standing up infrastructure along the way.」
Baseten 嘅 Applied Researcher Aaron Ellis-Bloor 補充:「smithtune's integration with Baseten Loops makes it seamless for teams to go from data collection to running fine-tuning experiments in minutes.」
官方話「A model trained on your examples can often perform as well as or better than a general purpose frontier model on specified tasks, often at lower cost and latency.」
邊個先玩到
官方冇公佈定價。
阿調搞兩下
呢條線路好清晰:你生產環境上跑緊嘅 agent,每一條 trajectory 都係「老師」喺成功情境下嘅示範;smithtune 就做嘅就係將呢啲示範直接「餵」俾學生模型。問題係邊個決定「教師數據」夠唔夠乾淨——LangSmith 話 trajectory format 為 post-training 而設,但數據選擇(data selection)究竟係靠人工篩定係有自動化機制,官方呢篇冇講清楚。另外「lower cost and latency」呢句係廠商聲稱,冇第三方跑分,亦冇樣本量同測試條件。對打工仔嚟講,呢個工具唔會幫你保住份工,但如果你公司正喺度揾方法慳 inference cost,呢個係值得追嘅產品。
阿調|深度調教




