Kimi K2.6評測：號稱千機大軍，官方文件只有300個 | Kimi K2.6 Review: 1,000-Agent Claims vs. 300 on Paper

10月 11, 2026

By Kit 小克 | AI Tool Observer | 2026-10-12

🇹🇼 Kimi K2.6評測：號稱千機大軍，官方文件只有300個

Kimi K2.6 Agent Swarm是什麼？中國開源模型打「群體戰」

這幾天開源 AI 圈最熱的不是哪家又刷新跑分，而是月之暗面（Moonshot AI）開源模型 Kimi K2.6 的 Agent Swarm（智能體群） 功能更新：多篇科技媒體報導稱它能同時調度上千個子代理（sub-agent）分工處理複雜任務，甚至示範用 10 小時造出一套 SysY 編譯器——這種活通常要一個團隊做好幾個月。但認真回頭查 Moonshot 官方文件與 Hugging Face 模型卡，數字沒那麼誇張。這正是這篇文章想拆給你看的地方。

官方規格：最多 300 個子代理，不是千個

Kimi K2.6 是 1 兆參數的 MoE 模型（實際啟用 320 億），官方 Agent Swarm 文件寫得很明確：最多可橫向擴展到 300 個子代理、執行超過 4,000 個協同步驟，比單一代理循序執行快約 4.5 倍。從 K2.5 的 100 個上限擴充到 300 個，確實是進步，但跟「上千個」的說法對不上——這種落差多半來自不同媒體引用了不同時間點的 demo 數字，沒有回頭對照官方技術文件。

跑分表現：開源模型裡的第一名

SWE-Bench Pro：58.6%，高於 GPT-5.4（57.7%）、Claude Opus 4.6（53.4%）、Gemini 3.1 Pro（54.2%）
Terminal-Bench 2.0：66.7%（K2.5 為 50.8%）
支援 12 小時以上連續編碼任務，不中斷執行
Artificial Analysis 排名：346 個模型中第 4，開源模型裡第 1

Claw Groups：人機協作還在摸索期

除了 Agent Swarm，Moonshot 也推出研究預覽功能 Claw Groups，讓人類與多個代理跨裝置共用同一個工作空間，概念像「真正的隊友」而非單向交辦。但目前連第三方測試者都指出，代理何時該把任務交回人類審核的觸發機制「尚未有文件說明」，還是個半成品。

定價與授權

官方 API 報價是每百萬輸入 token $0.95、輸出 $4.00；第三方平台如 OpenRouter 壓到輸入約 $0.43、輸出 $1.83，價差頗大。授權採開放權重（Modified MIT），可自行部署微調。

小克觀點：群體協作不是人人都需要

實際動手玩過的開發者回饋很實在：如果你的工作量沒有大到需要拆成 300 份平行處理，單一個代理反而更省事——協調成本本身就是成本。Agent Swarm 比較適合「一次性大型任務」（例如造編譯器、批量生成上百個頁面），不是日常小任務的標配。評估前先回頭看官方規格文件，別被媒體標題牽著走。

好不好用，試了才知道。

🇺🇸 Kimi K2.6 Review: 1,000-Agent Claims vs. 300 on Paper

What Is Kimi K2.6's Agent Swarm? China's Open-Source Model Goes Multi-Agent

The hottest open-source AI story this week isn't a new benchmark record — it's Moonshot AI's open-weight Kimi K2.6 and its updated Agent Swarm feature. Multiple outlets reported it can coordinate up to a thousand sub-agents on complex tasks, even showcasing a full SysY compiler built in 10 hours — work that normally takes a team months. Checking Moonshot's own documentation and Hugging Face model card tells a less dramatic story, and that's worth unpacking.

Official Spec: 300 Sub-Agents, Not 1,000

Kimi K2.6 is a 1-trillion-parameter MoE model (32B active). Moonshot's own Agent Swarm docs are explicit: it scales horizontally to 300 sub-agents executing over 4,000 coordinated steps, running roughly 4.5x faster than a single agent working sequentially. That's a real jump from K2.5's 100-agent cap, but it's nowhere near the "1,000 agents" framing that spread across secondary coverage — likely a case of outlets citing different demo numbers without checking the technical docs.

Benchmarks: Best Among Open-Weight Models

SWE-Bench Pro: 58.6%, ahead of GPT-5.4 (57.7%), Claude Opus 4.6 (53.4%), and Gemini 3.1 Pro (54.2%)
Terminal-Bench 2.0: 66.7%, up from K2.5's 50.8%
Supports 12+ hour uninterrupted coding runs
Ranked #4 of 346 models on Artificial Analysis, #1 among open-weight models

Claw Groups: Human-Agent Collaboration, Still Rough

Alongside Agent Swarm, Moonshot shipped a research preview called Claw Groups — a shared workspace where humans and multiple agents collaborate across devices as teammates rather than in a strict handoff chain. Even third-party testers note the trigger mechanism for when a task escalates back to a human "isn't documented yet." It's early.

Pricing and License

Moonshot's official API prices at $0.95 per million input tokens and $4.00 output; third-party hosts like OpenRouter undercut that significantly (around $0.43 input / $1.83 output). The model ships under a Modified MIT open-weight license, so self-hosting and fine-tuning are both options.

Kit's Take: Swarms Aren't for Everyone

Developers who've actually tried it report something refreshingly honest: if your workload doesn't need splitting into 300 parallel pieces, a single agent is simpler — coordination itself has a cost. Agent Swarm fits one-off, large-scale jobs (building a compiler, generating hundreds of pages at once), not routine daily tasks. Before you buy the headline, check the vendor's own spec sheet.

You won't know until you try it.

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

搜尋此網誌

AI小貼士