Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18, 2026: a model that accepts text, images, audio and video in a 1M-token context and is built to act on that input, not just describe it. It is available through Alibaba's Qianwen AI Platform as a hosted API, alongside a real-time variant, Qwen3.8-Omni-Flash-Realtime. The announcement describes API availability and does not announce a weights release, so this series treats it as an API model from a lab that also publishes open weights, not as an open-weight release.
The Qwen team's announcement frames the goal as moving omnimodal models from understanding content to planning tasks, calling tools and completing work, with audio and video as first-class inputs. The examples it gives are production-flavored: video editing, music video creation, film commentary, audio-visual summarization and real-time conversation.
The headline specifications, per the announcement:
The team also released supporting tooling: expanded Qwen-MM-Plugins for long-form audio and video workflows, and an open-sourced runtime called Qwen-Live Harness for continuous real-time interaction. The announcement describes the harness as open-sourced; it does not say the same of the model.
The announcement does not publish a simple per-token price table in US dollars. Instead it reports relative changes against the previous Qwen3.5-Omni-Plus: an API price per hour of audio input more than 98% lower, and per hour of audio-visual input more than 93% lower. Its methodology note says hourly figures are estimated as 30 times the cost of two minutes of source material, with video at 720p and one frame per second, and that text prices are quoted in Chinese yuan.
For buyers, that means two things. First, the hourly framing is the right unit for audio and video workloads, where token counts are hard to reason about. Second, the actual bill depends on resolution and frame rate settings, so a pilot on representative media is the only reliable estimate.
The announcement reports an average improvement of more than 25% over Qwen3.5-Omni-Plus across 29 audio, audio-visual and agent evaluations, and large gains on agent benchmarks such as WildClawBench-MM (36.5 points) and AgenticVBench (22.3 points). One striking example is meeting transcription: on the AliMeeting test, the team reports diarization error rate falling from 88.11 to 3.35. It describes audio-visual performance as close to Google's Gemini 3.8 Flash and overall audio performance as exceeding it.
These are the Qwen team's measurements, and this article does not reproduce them. The evaluation set includes Chinese-language speech benchmarks such as AliMeeting, AISHELL-4 and WenetSpeech, which is worth weighing for English-first deployments.
Teams building audio- and video-centric agents, such as meeting summarization, media production pipelines or real-time voice assistants that need to see and hear, have few models designed for that combination.
Teams already using Qwen open-weight models should note that this is a different product line. For open weights from the same lab, see Qwen-Image-2.1.
Teams comparing speech options can set it against Grok Voice Transcribe 2.0 for pure transcription workloads.
When was Qwen3.8-Omni-Flash released?
September 18, 2026.
Is Qwen3.8-Omni-Flash open-weight?
Not as announced. The Qwen team's announcement offers the model through the Qianwen AI Platform API and describes open-sourcing a companion runtime, Qwen-Live Harness, but does not announce a release of the model weights.
What inputs does Qwen3.8-Omni-Flash accept?
Text, images, audio and video, with a 1M-token context window.
Discover more content:
