Qwen3.8-Omni-Flash

The Qwen3.8 generation of Alibaba's omni-modal line, served as qwen3.8-omni-flash with no weights released. Qwen Cloud describes it as a "next-generation native omni-modal model" supporting 1M-token context that "natively accepts text, image, audio, and video inputs" and is built on the Qwen3.8-Flash-Next architecture. Output is text only — Qwen3.5-Omni remains the model to call for speech output — and the positioning has moved from understanding to acting: the product page frames it around agentic work in video editing, music-video creation, film production and narration, multimedia summarization and audio-video dialogue, and it supports thinking, function calling, structured outputs, prefix completion, context caching, batching and web search (the only built-in Responses tool so far). It also handles two- and four-channel spatial audio, and Alibaba recommends pairing it with the companion Qwen-MM-Plugins so agent frameworks can reach its native multimodal capabilities.

Documented limits are considerably wider than the previous omni tiers: audio input covers 113 languages and dialects, a single request may carry up to 64 files, 2 GB each by public URL, and up to 2 hours of audio or video (against 1 hour for the Qwen3.5-Omni series, 150 seconds for Qwen3-Omni-Flash and 40 seconds for the retired Qwen-Omni-Turbo, whose docs now tell text-analysis users to migrate here). Audio is billed at 7 tokens per second of input, against 12.5 for Qwen3-Omni-Flash and 25 for Qwen-Omni-Turbo. Qwen Cloud lists $0.15 per M input and $0.47 per M output tokens, $0.016 per M on implicit cache reads, with a 991K max input, 131K max output and a 262K reasoning budget; Model Studio serves it in Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia. Parameter count, training data and benchmark results are undisclosed, and Artificial Analysis had not scored it as of 2026-09-30. Alibaba's own pages carry no release date; the date here follows Chinese coverage published 2026-09-19 that puts the launch on 2026-09-18. The model reached OpenRouter on 2026-09-21.