Qwen3.8-Omni-Flash: One Model That Sees, Hears, and Calls Tools for Agents

Alibaba’s Qwen team is betting that the next coding-agent interface isn’t a chat box — it’s a meeting recording. Qwen3.8-Omni-Flash, released September 18 via Alibaba’s Qianwen AI Platform, accepts text, images, audio, and video through a single API call, with a 1-million-token context window and a rate card built for repeated media work rather than one-shot queries.

The practical shape of the model: one call can carry all four input types through the standard Chat Completions and Responses APIs. Native output is text — when the model needs to produce media, it goes through function-calling to external tools rather than generating it directly. Around it, Alibaba open-sourced two companion projects: Qwen-MM-Plugins (Video2Note, Music2MV, dubbing, commentary) and the Qwen-Live Harness for continuous audiovisual interaction, plus a realtime variant served over WebSocket with roughly 0.6 seconds to first token.

Priced for agents, not demos

The rate card is the story. International pricing lists $0.15 per million input tokens and $0.47 per million output tokens, with implicit cache hits at $0.016 per million — about a tenth of the standard rate, which matters for audiovisual agents that re-run requests over the same long recording. Alibaba claims audio input costs more than 98% less per hour than the previous generation and audio-visual input more than 93% less. The limits are concrete: video files up to 2 hours and 2 GB by URL, audio files up to 3 hours, audio input in 113 languages and dialects, and availability in six regions — Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia.

Separate the claims from the facts

Everything about relative quality is vendor-reported, so treat it accordingly. Alibaba claims an average improvement of more than 25% over Qwen3.5-Omni-Plus across 29 evaluations, and says its audio performance exceeded Gemini 3.8 Flash overall while audiovisual came close — selected, company-reported results, not independent benchmarks. Alibaba also hasn’t published the parameter count, training-compute budget, or training-data composition for this model, so developers get firm API specs and pricing with much less visibility into what’s underneath.

Note the naming trap: this is not the previously covered Qwen3.8-Flash (the multimodal MoE coding model) — Omni-Flash is the separate omni-modal line aimed at agentic media understanding. For teams building agents that watch, listen, and act — transcribing multi-hour meetings, analyzing video, dubbing content — the pitch is a single model ID at fifteen cents a million tokens. The independent benchmarks haven’t arrived yet. The API is live today.