Qwen3.8-Omni-Flash Release: Omnimodal Agentic Model with 1M-Token Context

TL;DR

Qwen3.8-Omni-Flash is a new omnimodal language model that supports text, image, audio, and video inputs with a 1 M‑token context window, delivers agentic planning and tool use for real‑world productivity workflows, and reduces audio‑visual API pricing by over 90% compared with its predecessor.

Core Technical Advances

1. 1 M‑Token Context Window – The model can ingest up to one million tokens of multimodal data while preserving text‑only performance comparable to a same‑size text‑only model.

2. Agentic Omnimodal Reasoning – Qwen3.8-Omni-Flash can start from a user question, plan a sequence of perception and tool‑use steps, and iteratively gather evidence from long audio‑visual streams. This agentic mode improves accuracy on OmniVideoBench from 63.4 % to 67.8 % while cutting token consumption by ~45.7 %.

3. Cost Reductions – The hourly price for audio input drops >98 % and for audio‑visual input >93 % relative to Qwen3.5-Omni-Plus, based on a standardized 720p @ 1 fps pricing methodology.

4. Scaling Gains – Across 29 benchmark evaluations the model’s average score rises >25 % over Qwen3.5-Omni-Plus, with especially large gains on WildClawBench‑MM (+36.5 points) and AgenticVBench (+22.3 points). Audio‑visual performance approaches or exceeds Gemini 3.8 Flash.

Benchmark Highlights

Benchmark Qwen3.8‑Omni‑Flash Qwen3.5‑Omni‑Plus Gemini 3.8 Flash
WildClawBench‑MM (multimodal tool use) 71.0 34.5 58.9
UniClawBench (multimodal tool use) 69.6 67.1 69.0
AgenticVBench (multimodal tool use) 36.8 14.5 45.0
OmniVideoBench (audio‑visual reasoning) – Static 63.4 53.8 65.2
OmniVideoBench – Agentic 67.8 53.8 65.2
OmniCap‑IF CSR 80.6 (↑8.5) 72.1 81.9
OmniCap‑IF ISR 28.2 (↑14.1) 14.1 28.3
AliMeeting DER / cpWER **3.4 17.2** (down from 88.1 89.6)

Scores in bold are the best among the listed systems. Improvements over Qwen3.5‑Omni‑Plus are shown in parentheses.

New Agentic Capabilities

Controllable Audio‑Visual Captioning

The model can generate video captions tailored to user‑specified subjects, time ranges, detail levels, and output formats. This enables use cases ranging from high‑level overviews to fine‑grained analyses of lighting, camera shots, and sound design.

Long‑Form Audio‑Visual Understanding

For multi‑hour recordings, the agent does not process the entire stream. Instead, it performs coarse‑to‑fine evidence gathering, focusing compute on segments that answer the query. Token usage is reduced by nearly half while accuracy improves.

Meeting Intelligence

Qwen3.8-Omni-Flash jointly diarizes speakers across audio and video, transcribes content, aligns identities, and can generate minutes, action items, and risk analyses. Integrated tool calls allow the agent to send follow‑up emails, create tasks, or even start coding based on meeting outcomes.

Research‑Assisted Video Exploration

When a user asks a question that extends beyond the video, the model can retrieve complementary multimodal sources (images, documents, other videos) and produce a richly illustrated research report, e.g., explaining Photoshop blend‑mode nuances.

End‑to‑End Audio‑Visual Production

The model supports full production pipelines such as music‑video creation, short‑drama localization, and long‑form film commentary. Users provide a single high‑level instruction and the agent plans, calls creative tools, and delivers a finished video.

Ecosystem Extensions

Qwen‑MM‑Plugins – A plugin suite that adds perception, tool use, and workflow execution for long‑form audio‑visual content. Plugins include:

  • omni-chatcut for music‑video generation
  • omni-video2note for turning tutorials into PDF notes
  • omni-skill-creator for extracting reusable SOPs from demonstrations
  • omni-memory for building persistent audio‑visual memory across sessions

Qwen‑Live Harness – An open‑source runtime that connects the realtime API (qwen3.8-omni-flash-realtime) to agent frameworks, handling continuous audio‑visual streams, memory management, and proactive task delegation.

Real‑Time Interaction (Qwen3.8‑Omni‑Flash‑Realtime)

The realtime variant processes live audio‑visual streams at ~85 tokens/s, with first‑token latency under 1 s for both text and audio. It supports:

  • Speaking practice – Aligns non‑standard accents to standard pronunciation while preserving rhythm and emotion.
  • Spatial audio perception – Locates sound sources in 3‑D space using visual context, enabling commands like “come over here”.
  • Dynamic knowledge injection – Skills can load brand language, business rules, or domain knowledge on‑the‑fly, allowing a single model to assume multiple roles.

API Usage Highlights

  • Supports OpenAI‑compatible chat completions.
  • reasoning_effort parameter (xhigh, medium, low) controls depth vs. cost.
  • preserve_thinking is enabled by default for transparent chain‑of‑thought output.
  • Example Python snippet demonstrates mixed‑modality input (image + audio + text) and streaming response.

Implications for AI‑Powered Productivity

Qwen3.8-Omni-Flash moves omnimodal models from passive perception to autonomous task execution. By integrating long‑form audio‑visual reasoning, agentic planning, and low‑cost APIs, it enables new classes of applications:

  • Fully automated video editing and localization pipelines.
  • Real‑time meeting assistants that not only summarize but also act on decisions.
  • Knowledge workers who can query hours of multimedia content with a single prompt and receive concise, actionable outputs.
  • Model‑driven R&D loops where a large model iteratively improves smaller specialist models, as demonstrated by the 40 % relative CER reduction on Sichuan‑dialect speech recognition.

Getting Started

  1. Obtain an API key from the Qianwen AI Platform.
  2. Use the OpenAI‑compatible endpoint with model name qwen3.8-omni-flash (or qwen3.8-omni-flash-realtime for live streams).
  3. Install the Qwen‑MM‑Plugins suite via the provided installer script to enable perception and tool‑use capabilities.
  4. For realtime applications, install websocket-client, pyaudio, and opencv-python and follow the sample WebSocket client code.

Citation

@misc{qwen38omniflash,
  title = {Qwen3.8-Omni-Flash: Omni Senses. Agentic Delivery.},
  url = {https://qwen.ai/blog?id=qwen3.8-omni-flash},
  author = {{Qwen Team}},
  month = {September},
  year = {2026}
}

Sources