Alibaba Qwen Releases Qwen3.8-Omni-Flash, a 1M-Context Agentic Omni-Modal Model

The model is hosted. Qwen said 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.

Qwen3.8-Omni-Flash is built on the Qwen3.8-Flash-Next architecture, which shipped with open weights in August 2026. The context window is 1 million tokens. QwenCloud lists 991,000 maximum input tokens and 131,000 maximum output tokens. Maximum reasoning length is 262,000 tokens. Output is text only; Model Studio documentation points developers to Qwen3.5-Omni when they need generated speech. Thinking is on by default with reasoning_effort set to xhigh, and setting it to none disables thinking. The API follows both the DashScope and OpenAI protocols and works with Chat Completions and the Responses API. Function calling, web search, structured outputs, context caching, and batch calls are supported.

For long video, Qwen described an agentic perception path. Instead of reading a long file from start to finish, the agent starts from the question, decides what to watch and hear, and gathers evidence over several coarse-to-fine rounds, sending compute and tokens to the segments that matter. On OmniVideoBench, Qwen reported accuracy rising from 63.4 to 67.8 while token use dropped from 145,736 to 79,117, about 45.7% fewer tokens.

All benchmark figures in the release came from Qwen, and independent results were not available at publication. Across 29 evaluations, Qwen said the average score improved by more than 25% over Qwen3.5-Omni-Plus. WildClawBench-MM improved by 36.5 points, AgenticVBench by 22.3 points, and LongAudioSpan by 8.3 points. UniClawBench reached 69.6, OmniVideoBench gained 9.6 points, and OmniCap-IF CSR and ISR improved by 8.5 and 14.1 points. The research team said audio-visual performance was close to Gemini 3.8 Flash and claimed overall audio performance above Gemini 3.8 Flash. A Qwen post on X summarized the agent gains as an average of 19.5 points across WildClawBench-MM and UniClawBench.

QwenCloud lists pricing at $0.15 per 1 million input tokens and $0.47 per 1 million output tokens. Implicit cache hits cost $0.016 per 1 million tokens. Qwen reported large cost cuts against Qwen3.5-Omni-Plus: audio input costs more than 98% less per hour, audio-visual input costs more than 93% less per hour, and the X post put the video input reduction at about 89%. Model Studio documentation lists limits: video files up to 2 hours and 2 GB by URL, audio files up to 3 hours, audio input in 113 languages and dialects, stable results with video sampled at up to 15 frames per second, and two-channel stereo and four-channel FOA spatial audio through use_multichannel. The Model Studio documentation lists availability in six regions: Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia.

Because the model returns text, Qwen said tools do the media work. The team is open-sourcing two projects, Qwen-MM-Plugins and a Qwen-Live harness. Qwen-MM-Plugins is live under Apache-2.0 with the tagline "Make any agent harness multimodal-native." Each capability installs as a Skill plus an optional MCP server, and the guided installer supports Claude Code, CodeBuddy, Codex, Qoder, OpenClaw, Qwen Code, and Gemini CLI. The omni capabilities map to launch demos: omni-memory builds an audio-visual memory of a long video, omni-video2note converts a tutorial video into an illustrated PDF, and omni-chatcut covers Music-to-MV, movie commentary, and speaker-preserving video translation. A core plugin lets the main model read local images and video frames natively.