AWS Strands Labs Releases Strands Decider 2B: An Open Source Decision Model That Picks Options in About 115 ms
https://www.marktechpost.com/2026/10/01/aws-strands-labs-releases-strands-decider-2b/📌 AWS 開源 Strands Decider 2B：一次前向傳播、115 毫秒選出答案的決策模型
TL;DR：AWS Strands Labs 推出開源「決策模型」，不生成文字只選選項，19 億參數可在本機跑。
如果你的 agent 只是要在「billing / sales / retail」之間選一個分類，為什麼要讓一個會寫詩、會聊天、會生成任意文字的 LLM 跑一整輪 decoding？AWS Strands Labs 給出的答案是：換一顆完全不同設計的模型。
🤔 System One 模型賽道，再添一名玩家
這類「決策模型」（decision model）也被稱為 System One 模型，是在 TypeSafe AI 上個月推出 Jev 之後才成形的新類別。概念很單純：LLM 能生成任意輸出，而決策模型只在固定選項中選一個，或者在量表上打分。AWS Strands Labs 這次推出的 Strands Decider 2B，就是這條路線上的開源實作。它讀取一段 state 與一組 typed questions，回傳的是一個選項、一個 yes/no 機率，或一個帶校準信心值（calibrated confidence）的分數，答案一律來自請求裡給定的選項集合。
🧩 拿掉語言模型頭，換上一個百萬參數的指標頭
Strands Decider 從 Qwen3.5-2B-Base 出發,先把語言建模頭（language-modelling head）整個拿掉,換上一個約 100 萬參數的小型 pointer head。這個頭的做法是比對
📊 JevBench 排名第 3,但校準才是真正賣點
團隊在第三方基準 JevBench 上測準確率與校準度。在 9 月 25 日的 v1.4.2 榜單上,v19 在 2B 量級中排名第 3 of 33;若排除 3 個略超過 2B 的模型,則是第 1 of 30。不過倉庫裡也標註了一個對照:Mapika 較新的 decider-2b v11 在 Strands harness 上拿到 175/231,領先 8 個任務;而 Strands Decider 在撰文當時尚未出現在更新的 v1.5.4 綜合榜上。團隊特別強調,校準才是實務上真正有價值的部分——在未見過的短文字分類任務上,信心值在 0.9 以上的答案,正確率約達 95%,團隊建議信心低於這個門檻時就該確認或升級處理。需要留意的是,各項延遲數字來自不同硬體與測試框架,彼此並不能直接比較。
💡 路由、守門、混合式 agent 是主要落地場景
團隊回報的早期成功案例集中在模型路由、工具選擇、參數檢查、分流（triage）、guardrails、評測以及混合式 agent。在混合式 agent 的設計裡,LLM 負責處理困難判斷,decider 負責處理瑣碎但頻繁的例行決策。倉庫範例示範了在 Strands agent 裡用 before_tool_call 攔截天氣工具呼叫,問兩個 yes/no 問題:參數是否確實根據使用者說的話、現在呼叫是否過早;如果 agent 是猜測城市,就會改成反問使用者。CLI 範例裡,把「Help! My payouts have been failing for 3 days!」這句話路由到 billing / sales / retail returns 三個分類,結果選中 billing,信心值 0.768。
⚠️ 複雜任務別指望它,預設也沒做好生產環境防護
團隊自己承認,這顆模型在複雜問題上不如推理模型,完全不適合用來寫程式、聊天或做摘要。另外,內建的 HTTP server 預設綁定在 127.0.0.1 且沒有任何驗證機制,要上生產環境必須自己加一層身份驗證;目前也還沒有任何雲端推理供應商託管這顆模型。
🎯 實務啟示
如果你的 agent pipeline 裡有大量「選選項、打分數、做二元判斷」的瑣碎決策,把這些交給一顆 19 億參數、CPU 就能跑、單次前向傳播即可出結果的小模型,很可能比每次都呼叫 LLM 更划算;但校準門檻與升級路徑要自己設計好,別指望它處理灰色地帶。
🔗 來源
- 標題：AWS Strands Labs Releases Strands Decider 2B: An Open Source Decision Model That Picks Options in About 115 ms
- 作者／機構：Asif Razzaq, MarkTechPost
- 連結：https://www.marktechpost.com/2026/10/01/aws-strands-labs-releases-strands-decider-2b/
#DecisionModel #AWS #OpenSource #SystemOneAI #AgentRouting #LLM #MachineLearning #Qwen #Calibration #AIInfrastructure
原始資料 MarkTechPost · 收集於 2026-10-02
摘要原文
AWS Strands Labs releases Strands Decider 2B , an open source decision model. It does not generate text. It reads a state and typed questions, then returns a choice, a yes/no probability, or a score with a calibrated confidence. The model has 1.9 billion parameters and runs locally on a CPU, a consumer GPU, or an Apple silicon Mac. Is it deployable? Yes, for local and self-hosted use. Weights are on Hugging Face under Apache-2.0, and pip install strands-decider gives a CLI and an HTTP server. The bundled server binds to 127.0.0.1 with no authentication, so production needs your own auth layer. No hosted inference provider serves it yet. Decision models, also called System One models, became a category after TypeSafe AI launched Jev last month. An LLM can produce arbitrary output. A decision model only picks between options or rates on a scale. Strands Decider supports 3 question types: Every answer comes from the allowed options and carries a confidence. The team states the model is worse than reasoning models on complex problems. It is unsuited for coding, chat, or summarization. The team starts from Qwen3.5-2B-Base and discards the language-modelling head. A small pointer head of about 1 million parameters replaces it. That head compares the hidden state at the <answer> position against the hidden state at each option’s last token. One forward pass yields the result, with no decoding loop. The torso uses a rank-16 LoRA, and the head runs in fp32. Label sets come from the request, so nothing caps the option count. The released checkpoint is v19. Asking several questions about one text is cheap. The state is read once, and each extra question adds only its own tokens. The team measures accuracy and calibration on the public set of JevBench , a third-party benchmark for Jev-class models. Published v19 figures: On the v1.4.2 board of September 25, v19 ranked 3rd of 33 in the 2B class. Excluding 3 models just over 2B, it ranked 1st of 30. The repo also flags a caveat. Mapika’s newer decider-2b v11 scores 175 of 231 on the Strands harness, 8 tasks ahead. Strands Decider was not on the newer v1.5.4 composite board at the time of writing. Calibration is the practical win. On unseen short classification tasks, answers at 0.9 confidence or higher were right about 95% of the time. The team advises confirming or escalating below that threshold. Sources: Strands JevBench comparison , Benchmark Heaven JevBench , Jev product page . Latencies come from different hardware and harnesses, so they are not directly comparable. The team reports early success in model routing, tool selection, argument checking, triage, guardrails, evals, and hybrid agents. In a hybrid agent, an LLM makes the hard calls and the decider handles rote ones. The repo example gates a weather tool call inside a Strands agent. A before_tool_call intervention asks 2 yes/no questions. Are the arguments grounded in what the user said? Is calling now premature? If the agent guessed a city, it asks the user instead. From the CLI, routing “Help! My payouts have been failing for 3 days!” across billing, sales and retail returns billing with confidence 0.768. Check out the technical details , GitHub repo and model weights . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post AWS Strands Labs Releases Strands Decider 2B: An Open Source Decision Model That Picks Options in About 115 ms appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成