Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use
https://www.marktechpost.com/2026/09/18/alibaba-qwen-releases-qwen3-8-omni-flash/📌 Qwen3.8-Omni-Flash：百萬 context 全模態模型的 agentic 賭注
TL;DR：阿里 Qwen 推出首個強調 agentic 能力的全模態模型，僅開放 API，不開權重。
多數影片模型的做法是把整段影片從頭看到尾，即使答案只藏在 3 分鐘的片段裡，也照樣把全部 token 吃下去。Qwen 團隊這次選擇了不同的路：讓模型先想清楚問題,再決定該看哪裡、聽哪裡。
🤔 從「全部看完」到「先問問題再找證據」
Qwen3.8-Omni-Flash 是 Qwen 團隊發布的首個「built around agentic capabilities」的全模態模型，可接受文字、圖片、音訊、影片輸入，輸出則僅限文字。官方描述的工作流程是：理解內容 → 規劃任務 → 用工具執行 → 交付結果。針對長影片，模型會以問題為起點，決定要看什麼、聽什麼，再透過多輪由粗到細（coarse-to-fine）的方式蒐集證據，把運算資源和 token 集中花在真正相關的片段上。
🧩 建構在 Qwen3.8-Flash-Next 之上，僅供 API 呼叫
這個模型是以 2026 年 8 月開源的 Qwen3.8-Flash-Next 架構為基礎打造，但發布時並未釋出開放權重，目前只能透過 QwenCloud、Alibaba Cloud Model Studio、Qwen Studio 以 hosted API 形式使用，無法自架部署。
context window 為 1M tokens，QwenCloud 文件標示最大輸入 991K、最大輸出 131K，最大 reasoning 長度 262K tokens。thinking 預設開啟（reasoning_effort 設為 xhigh），設為 none 可關閉。若需要語音生成，Model Studio 文件會導向 Qwen3.5-Omni，因為這款模型只輸出文字。API 同時相容 DashScope 與 OpenAI 協定，支援 Chat Completions 與 Responses API，並具備 function calling、web search、structured outputs、context caching、batch calls。
📊 官方數據：token 省下近半，準確率不降反升
Qwen 團隊公布的 OmniVideoBench 結果如下（皆為官方數字，發文時尚無第三方驗證）：
| 指標 | 數值 |
|---|---|
| 準確率 | 63.4 → 67.8 |
| Token 用量 | 145,736 → 79,117（約減少 45.7%） |
Qwen 團隊另表示，其 audio-visual 表現接近 Gemini 3.8 Flash，整體 audio 表現則宣稱優於 Gemini 3.8 Flash；官方 X 貼文則將 agent 能力提升總結為在 WildClawBench-MM 與 UniClawBench 上平均 +19.5 分。
定價方面，QwenCloud 列出每 1M input tokens $0.15、每 1M output tokens $0.47，implicit cache 命中則為每 1M tokens $0.016。相較於 Qwen3.5-Omni-Plus，官方宣稱 audio 輸入成本每小時降低超過 98%，audio-visual 輸入降低超過 93%；X 貼文另提及影片輸入成本降低約 89%。
💡 兩個開源專案補上「多模態原生」的最後一哩
由於模型本身只回傳文字，多媒體相關操作需仰賴外部工具，Qwen 團隊同步開源了 Qwen-MM-Plugins（Apache-2.0 授權），標語是「Make any agent harness multimodal-native」。每項能力以 Skill 加上可選的 MCP server 形式安裝，並提供 guided installer，支援 Claude Code、CodeBuddy、Codex、Qoder、OpenClaw、Qwen Code、Gemini CLI 等多種 agent harness。其中一個核心 plugin 可讓主模型原生讀取本地圖片與影片幀。
⚠️ 仍有明顯限制
README 也坦承現況的一個缺口：目前多數 agent harness 還無法讓主模型原生接收音訊，音訊處理暫時仍得透過 API 轉送。此外，模型未開放權重意味著無法自架，且輸出僅限文字，語音生成需另外導向 Qwen3.5-Omni。
🎯 對工程師的意義
如果你的應用需要處理長影片或音訊並串接工具鏈，Qwen3.8-Omni-Flash 的 coarse-to-fine 檢索策略與大幅降低的 token 成本值得評估；但若你的專案要求自架部署或資料不出雲端，目前這條路是走不通的，只能等待未來是否釋出開放權重版本。
🔗 來源
- 標題：Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use
- 作者／機構：Asif Razzaq, MarkTechPost
- 連結：https://www.marktechpost.com/2026/09/18/alibaba-qwen-releases-qwen3-8-omni-flash/
#Qwen #Alibaba #OmniModal #LLM #AIAgents #MultimodalAI #VideoUnderstanding #APIModel #AIInfrastructure #ToolUse
原始資料 MarkTechPost · 收集於 2026-09-18
摘要原文
Alibaba’s Qwen team has released Qwen3.8-Omni-Flash . They called it its first omni-modal model built around agentic capabilities. It accepts text, images, audio, and video, and it returns text. Audio-video understanding, reasoning, and tool use sit inside one model. The stated workflow is simple: understand the content, plan the task, execute with tools, deliver the result. Is it deployable? Yes, as a hosted API today. It is live on QwenCloud , Alibaba Cloud Model Studio , and Qwen Studio . No open weights were announced at launch, so self-hosting is not an option. The Qwen3.8-Omni-Flash is built on the Qwen3.8-Flash-Next architecture. That base model shipped with open weights in August 2026. The context window is 1M tokens. QwenCloud lists 991K max input and 131K max output. Max reasoning length is 262K tokens. Output is text only. The Model Studio docs point developers to Qwen3.5-Omni when they need generated speech. Thinking is on by default, with reasoning_effort set to xhigh . Setting it to none disables thinking. The API follows both the DashScope and OpenAI protocols. It works with Chat Completions and the Responses API. Function calling, web search, structured outputs, context caching, and batch calls are supported. Most video models read a long file from start to finish. That holds even when the answer sits in 3 minutes of footage. Qwen research team describes a different path. The agent starts from the question. It decides what to watch and hear. It then gathers evidence over several coarse-to-fine rounds. Compute and tokens go to the segments that matter. The research team reports the result on OmniVideoBench. Accuracy rises from 63.4 to 67.8. Token use drops from 145,736 to 79,117. That is about 45.7% fewer tokens. All figures here come from Qwen. Independent results were not available at publication. The research team states that audio-visual performance is close to Gemini 3.8 Flash. It also claims overall audio performance above Gemini 3.8 Flash. The X post summarizes the agent gains as +19.5 points on average across WildClawBench-MM and UniClawBench. Meet Qwen3.8-Omni-Flash, Qwen's first omni-modal model built around agentic capabilities! Native audio-video understanding, reasoning, and tool use come together in one model: understand the content, plan the task, execute with tools, and deliver the result. Highlights: -… pic.twitter.com/iJypeohw7y QwenCloud lists $0.15 per 1M input tokens and $0.47 per 1M output tokens. Implicit cache hits cost $0.016 per 1M tokens. The research team reports large cost cuts against Qwen3.5-Omni-Plus. Audio input costs over 98% less per hour. Audio-visual input costs over 93% less per hour. The X post puts the video input reduction at about 89%. Key limits from the Model Studio docs : Calling it takes a few lines with the OpenAI SDK: The model returns text, so tools do the media work. Qwen team is open-sourcing 2 projects to support that. Qwen-MM-Plugins is live under Apache-2.0. Its tagline is ‘Make any agent harness multimodal-native.’ Each capability installs as a Skill plus an optional MCP server. The guided installer supports Claude Code, CodeBuddy, Codex, Qoder, OpenClaw, Qwen Code, and Gemini CLI. The Omni capabilities map to the launch demos: A core plugin lets the main model read local images and video frames natively. The README notes one current gap. Most harnesses cannot feed audio to the main model natively yet. Audio is routed through the API for now. Check out the Technical details , QwenCloud model page , API docs , and GitHub repo . 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 Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成