Alibaba's Qwen team shipped Qwen3.8-Omni-Flash today. The number that got me: overlapping multi-speaker meetings used to be a diarization disaster, 88% error. Now 3.4%.
Rest of the package: native text/image/audio/video in, 1M context, +25% average over their own Qwen3.5-Omni-Plus across 29 tests, audio input pricing cut over 98%. There's a Realtime API too, first token under a second on most inputs. Plus two open toolkits. Qwen-MM-Plugins drops the model into Claude Code, Gemini CLI, Codex, OpenClaw. Qwen-Live-Harness does task delegation, the "ping me when the job's done" kind.
ngl omni-skill-creator is the sleeper. Record yourself operating software, narrate why as you go, and it turns the video into a reusable agent skill. Show once, the agent learns.
And it's already live on Vercel's AI Gateway, so the "open models are hard to reach from abroad" excuse keeps getting thinner. The gateway became the on-ramp.
Would you rather record one video and let it learn, or keep writing agent skills by hand?
https://www.threerouter.com/blog/57
FreyaLi
Posted on
Qwen3.8-Omni-Flash: Multimodal Model Released
For further actions, you may consider blocking this person and/or reporting abuse
Top comments (0)