
基於阿里巴巴 Qwen3.5-9B 進行後訓練,旨在提升自動化系統處理速度與成本效益
微軟於 Microsoft Foundry 發布 Decision-1 決策模型,專門處理分類與評分任務。該模型基於 Qwen3.5-9B 進行後訓練,不生成文字,而是輸出機率分數以供自動化系統決策,目前已在 OpenRouter 上架。
AI-generated summary
微軟於 10 月 9 日推出專門處理分類與評分任務的 Decision-1 模型,該模型不生成文字,而是輸出機率分數。
微軟10月9日在 Microsoft Foundry 推出 Microsoft-Decision-1 決策模型,專門處理分類、選項判斷與評分工作。這款模型不採用一般生成式 AI 的文字回答方式,而是根據輸入情境和預先提供的選項,計算各選項的機率分數,供應用程式決定下一步操作。第三方 AI 模型平台 OpenRouter 也同步上架,模型代號為「microsoft/microsoft-decision-1」。
Decision-1 的用途與一般聊天模型不同。開發者可以提供一段使用者訊息,要求模型判斷其意圖,再依照結果轉交客服、技術支援或其他處理流程;也能用於資料分類、事件分流、內容安全標記及量表評分。模型支援是非題、多選題,以及無法從現有資訊作出判斷時的棄權選項。由於輸出限定為決策分數,不需要另外生成文字解釋,適合整合進需要快速判斷的自動化系統。
微軟將 Decision-1 設計為純文字模型,上下文長度最高 32,768 token,不接受影像、音訊或影片。模型權重並未公開,開發者須透過託管 API 使用,計費為每百萬輸入 token 0.042 美元,輸出不另外收費。Microsoft Foundry 提供 Decision 類型 API,部署選項包括 GlobalStandard,以及部分區域適用的 DataZoneStandard。
在模型來源方面,Decision-1 並非完全由微軟自家模型從頭訓練,而是以阿里巴巴開放權重的 Qwen3.5-9B 為基礎,再由微軟進行後訓練。外媒《The Register》指出,微軟尚未解釋初期選用 Qwen 的原因。微軟則表示,後續計畫改以自家 MAI 或 OpenAI 模型作為底層架構。儘管 Decision-1 使用開放權重模型開發,微軟並未對外提供完成後訓練的模型權重。
微軟執行長納德拉(Satya Nadella)也在社群平台 X 介紹這款模型,表示公司內部已將相關技術用於事件回應、品質控管及科學發現。官方公布的測試涵蓋36項基準、147,137道未納入訓練的題目,平均準確率達83.5%,回應延遲中位數約85毫秒。
Xbox Research、Copilot 和 Discovery 等微軟內部團隊也回報,在維持相近品質的情況下,Decision-1 能提高處理速度並降低成本。不過,這些效能與成本數據均來自微軟自身測試,尚無獨立第三方驗證。不同任務的選項設計、輸入長度與判斷門檻,也可能影響實際部署結果。
決策模型市場目前也已有其他業者投入,OpenAI 推出 Decisions API 後,市場上已有超過百款類似模型。微軟這次以獨立 API 提供 Decision-1,讓開發者可以將決策評分與一般文字生成工作分開處理,例如先判斷請求類型,再決定是否交由其他模型或程式執行。
AI outlook — possibilities, not facts
微軟計畫將底層架構更換為自家 MAI 或 OpenAI 模型。
Likely · Within months
The 12th Beijing International Aging Industry Expo opened in Beijing. Many companies displayed elderly care robots, exoskeletons and physical therapy equipment. Experts point out that elderly care robots have great potential in disability care and emotional companionship, but they still need to be further improved in terms of technical stability and standard systems.

Communist Party officials in China's Zhejiang province are demanding increased oversight of major internet platforms, arguing that algorithms and traffic power now significantly influence public sentiment and should be managed as semi-public utilities.
On October 10, AutoNavi, a subsidiary of Alibaba Group, and Sands China held an appointment ceremony at the Venetian NBA House in Macao to officially deliver the first batch of embodied robot dogs "Amap Tutu". Titu will serve as Sands China’s 001 robot dog guide, providing tour guides and global indoor navigation services for tourists.
In the offline closed-loop test of Jingdong Health's self-developed large-scale medical model Jingyi Qianxun, it successfully passed ten subjects of the senior professional title examination using real test questions from real candidates, demonstrating its ability in complex case handling and long-chain clinical reasoning.

Companies and their booking systems such as Peach Airlines, BOOKOFF and JR East have been subject to improper access in recent days, resulting in the leakage or delayed delivery of millions of emails and personal data. The Japanese government has asked operators to quickly update software and strengthen multi-factor authentication.
The 2026 Wuhan Artificial Intelligence and Robotics Expo opened on the 10th. More than 200 companies participated in the exhibition with products, attracting purchasing groups from more than 20 countries. A variety of innovative robots and aircraft were unveiled and a large number of overseas intended orders were obtained, demonstrating Wuhan's advantages in the artificial intelligence and robotics industry.