Qwen3.8-Omni-Flash: 全模态 Agent 型模型 + MM-Plugins 开源
- ID: 3195ca01
- 原文链接: https://qwen.ai/blog?id=qwen3.8-omni-flash
- 作者: Qwen Team
- 日期: 2026-09-14
- 分类: models
- 来源类型: article
- 标签: qwen, omni, multimodal-agent, mm-plugins, model-release
- 质量评分: 5/5
- 抓取时间: 2026-09-19T22:35:00Z
中文导读
Qwen 首个以 agentic 能力为核心的原生全模态模型 Qwen3.8-Omni-Flash 上线，集本地理解/推理/工具调用于一身；以音频为中心的场景包括视频剪辑、音视频剪辑、音视频转图文摘要与对话。重点指标：WildClawBench-MM 上提升 36.5 分、AgenticVBench 上提升 22.3 分、UniClawBench 69.6；OmniGAIA 未使用 harness。1M token 上下文环境里 OmniVideoBench 以比静态理解少 51.8% 的 token 获得更高准确率；视频输入成本相比 Qwen3.5-Omni-Plus 下降约 89%。同步开源 Qwen-MM-Plugins（多模态 Skill + MCP 服务组件，已适配 Claude Code/Codex/OpenClaw/Qwen Code/Hermes Agent 等多个 harness）与 Qwen-Live Harness。边界：价格依赖国际控制台，同类 WildClawBench 为厂商自测；Live Harness 官方页还可能 404。
为什么值得关注
Qwen 全模态 agent 模型上线：WildClawBench +36.5 分，视频输入成本下降 89%，同步开源 MM-Plugins
关键信息
- 标题：Qwen3.8-Omni-Flash: 全模态 Agent 型模型 + MM-Plugins 开源
- URL：https://qwen.ai/blog?id=qwen3.8-omni-flash
- 发布时间：2026-09-14
- 分类：models
- 标签：qwen, omni, multimodal-agent, mm-plugins, model-release
English Summary
Qwen released Qwen3.8-Omni-Flash, its first omni-modal model built around agentic capabilities, packaging native audio-video understanding, reasoning, and tool use into one checkpoint. Key scores: +36.5 on WildClawBench-MM, +22.3 on AgenticVBench, and 69.6 on UniClawBench (OmniGAIA runs without a harness); on OmniVideoBench the 1M-token context reaches higher accuracy with 51.8% fewer tokens than static understanding, and video input cost is down ~89% versus Qwen3.5-Omni-Plus. The launch also ships Qwen-MM-Plugins (a multimodal Skill + MCP package; installer covers Claude Code, Codex, OpenClaw, Qwen Code, Hermes Agent and others) and Qwen-Live Harness. Boundaries: pricing follows the international Model Studio console; benchmark numbers come from the vendor's own evaluation; the Qwen-Live Harness official page may still 404.