Alibaba releases Qwen3.8-Omni-Flash, a hosted 1M-token omni model that plans what to watch in video
- Alibaba's Qwen team released Qwen3.8-Omni-Flash, a cloud-hosted model that takes text, images, audio and video as input and returns text only, with a context window near 1 million tokens (reported max input around 991,000) and support for 2-hour video or 3-hour audio in a single request.
- The model's headline feature is agentic perception: instead of sampling every frame at fixed intervals it decides what to examine, then gathers evidence in several coarse-to-fine passes. On an internal long-video benchmark that lifted accuracy from 63.4 to 67.8 while cutting tokens per query about 46%, from roughly 145,700 to 79,100.
- Internally it is a sparse mixture-of-experts model with about 125 billion total parameters, of which roughly 6 billion activate per token across 512 experts, built directly on the Qwen3.8-Flash-Next architecture whose open weights shipped in August 2026. This version has no open weights and runs only as a hosted service.
- Pricing is $0.15 per million input tokens and $0.47 per million output tokens, with cached input at $0.016; measured per hour of media, Alibaba says audio input costs fall about 98% and combined audio-video costs about 89% versus Qwen3.5-Omni-Plus. It is live in six regions, including Singapore, Tokyo and Frankfurt, through Qwen Chat, Alibaba Cloud Model Studio and the QwenCloud API.
- Alibaba also open-sourced Qwen-MM-Plugins under Apache 2.0 so coding-agent harnesses such as Claude Code, Codex, Gemini CLI and Qwen Code can read images, video frames and audio locally, and previewed a real-time streaming counterpart with no release date. Every benchmark and cost figure comes from Alibaba; no independent evaluation has appeared, and analysts say the model approaches Gemini 3.8 Flash on audio-visual tasks while Gemini still leads some pure video-reasoning benchmarks.