Qwen3.8-Omni-Flash

What is disclosed

Qwen3.8-Omni-Flash exists only on the API, as qwen3.8-omni-flash, and its weights are not published.¹¹1QwenCloud, Qwen3.8-Omni-Flash model page, prices per million tokens as listed on 29 September 2026. https://www.qwencloud.com/models/qwen3.8-omni-flash QwenCloud gives no parameter count and no expert count.¹¹

According to the model page, it is built on the Qwen3.8-Flash-Next architecture.¹¹ That open model, published as Qwen3.8-Flash, has 125B parameters with 6B active per token. Its layers alternate three Gated DeltaNet layers with one layer of Qwen Sparse Attention.²²2Qwen, Qwen3.8-Flash-Next model card, August 2026: 125B parameters with 6B active, hidden layout 12 x (3 x Gated DeltaNet, 1 x Qwen Sparse Attention). https://huggingface.co/Qwen/Qwen3.8-Flash-Next Qwen does not say whether Omni-Flash keeps those sizes.

Qwen calls the model natively omni-modal. Audio and video go in alongside text and images.¹¹ It takes up to an hour of audio and video as input, a limit the announcement ties to meetings.³³3Alibaba Cloud, "Qwen3.8-Omni-Flash: Omni Senses. Agentic Delivery.", 20 September 2026. Up to one hour of audio-visual input; price comparisons are per hour of input, against Qwen3.5-Omni-Plus. https://www.alibabacloud.com/blog/603580 Speech recognition covers 74 languages and 39 Chinese dialects, Cantonese among them.³³ The model also reads two-channel and four-channel spatial audio.¹¹

The context window is 1M tokens. Input is capped at 991K, or 983K with thinking on, and output at 131K.¹¹

What is new

Qwen announced the model on 18 September 2026.⁴⁴4Qwen, Qwen3.8-Omni-Flash blog post, 18 September 2026. https://qwen.ai/blog?id=qwen3.8-omni-flash Across 29 evaluations it reports an average more than 25% above Qwen3.5-Omni-Plus, its omni model of an earlier generation.³³

Against Qwen3.5-Omni-Plus, an hour of audio input costs more than 98% less. An hour of audio with video costs more than 93% less.³³

QwenCloud aims the model at agent work. Coding and GUI interaction are on its list, and so are video editing and dialogue over audio and video.¹¹

Using it

A million input tokens cost $0.15 and a million output tokens $0.47. Cached input costs $0.016 per million.¹¹ Qwen3.8-Flash, built on the same architecture, costs the same.⁵⁵5QwenCloud, Qwen3.8-Flash model page, as listed on 29 September 2026: $0.15 input and $0.47 output per million tokens. https://www.qwencloud.com/models/qwen3.8-flash

With thinking on, the same model id takes 8K fewer input tokens and keeps the 131K output limit.⁶⁶6QwenCloud, Qwen3.8-Omni-Flash model page. Arithmetic: 991K minus 983K = 8K. https://www.qwencloud.com/models/qwen3.8-omni-flash

The standard API returns text only.¹¹ Yet the announcement describes speech generation in 29 languages and 7 dialects, Southern Min among them.³³ Live audio and video streams go to a separate realtime version, Qwen3.8-Omni-Flash-Realtime, over WebSocket or WebRTC.³³