Alibaba's Qwen team on September 18 released Qwen3.8-Omni-Flash, a single model that ingests text, images, audio and video inside a one-million-token context window and is priced at $0.15 per million input tokens and $0.47 per million output tokens. It is, by Alibaba's own account, the first Qwen model built from the ground up for agentic work rather than perception alone. It is also something the Qwen line has rarely been: a flagship with no downloadable weights, available only through Alibaba's hosted APIs.
The company says the model improves its average score by more than 26 percent across roughly 30 evaluations compared with its predecessor, Qwen3.5-Omni-Plus, with the largest jumps on agent-style benchmarks. On WildClawBench-MM, a multimodal agent test, Alibaba reports a 36.5-point gain to 71.0; on AgenticVBench it reports a 22.3-point gain. The Qwen team also claims audio-visual performance close to Google's Gemini 3.8 Flash and audio performance that beats it, though every figure so far comes from Alibaba and no independent replication has been published. Coverage from Alibaba's own blog and from GIGAZINE puts the comparison at 25 percent across 29 benchmarks, so the headline number varies slightly depending on which of the company's materials is cited.
“Native audio-video understanding, reasoning, and tool use come together in one model: understand the content, plan the task, execute with tools, and deliver the result,” the Qwen team wrote in its launch post on X.
The most concrete efficiency claim concerns what Alibaba calls agentic perception. Rather than reading a two-hour video frame by frame, the model starts from the user's question and decides which segments to watch and listen to, gathering evidence in several coarse-to-fine passes. On OmniVideoBench, the team says this lifted accuracy from 63.4 to 67.8 while cutting token consumption from 145,736 to 79,117, a reduction of about 45.7 percent. Because the model is billed per token, that translates directly into cost: Alibaba says an hour of audio input now costs more than 98 percent less than on Qwen3.5-Omni-Plus, and an hour of mixed audio-video input costs more than 93 percent less. Cached input is priced at $0.016 per million tokens.
The specifications are aimed squarely at long-form media. QwenCloud lists 991,000 tokens of maximum input, 131,000 tokens of output and a reasoning budget of up to 262,000 tokens. Video files up to two hours and two gigabytes can be passed by URL, audio files up to three hours, with speech input covering 113 languages and dialects and video sampled at up to 15 frames per second. Output is text only; developers who need generated speech are pointed back to Qwen3.5-Omni. The API supports function calling, web search, structured outputs, context caching and batch jobs, and is served from six regions including Singapore, Frankfurt and Virginia.
In its technical write-up, Alibaba framed the release as a shift in what audio and video are for. “These advancements mean that audio and video are evolving from perceptual inputs to core media for agents to understand and reason about their environment and perform tasks,” the company wrote. To back that up, it open-sourced Qwen-MM-Plugins under Apache 2.0, a set of skills and MCP servers that let harnesses such as Claude Code, Codex, Gemini CLI and Qwen Code feed long video into the model for tasks like turning a tutorial into an illustrated PDF or cutting a music video. A second project, Qwen-Live Harness, targets the real-time API, though its GitHub page was returning a 404 error at launch.
What Alibaba did not open is the model itself. Qwen3.8-Omni-Flash is built on the Qwen3.8-Flash-Next architecture, whose base weights shipped in August, but the omnimodal version is confined to QwenCloud, Alibaba Cloud Model Studio and Qwen Studio. Self-hosting is not an option.
That decision lands alongside a second, quieter signal. Over the same weekend the Qwen team put Qwen-Image-2.1 on Hugging Face and ModelScope: a unified generation-and-editing model with a 7-billion-parameter diffusion transformer, native 2048-by-2048 output, RGBA transparency support and up to ten reference images per prompt. The weights are public and the model runs on a consumer GPU. But the earlier Qwen-Image line shipped under Apache 2.0, and this one ships under a new Qwen Research License Agreement, dated September 20, that grants rights “for non-commercial purposes only” and routes any commercial use to a separately negotiated license with Alibaba.
Why It Matters
Qwen has spent two years as the most reliable source of permissively licensed frontier-adjacent models, and that reputation underwrote a large ecosystem of fine-tunes, quantizations and edge deployments. Two releases in one weekend now point the other way: the best omnimodal model is API-only, and the best image model is open-weight but closed for business. Neither is unprecedented on its own. Taken together, they suggest Alibaba is drawing a line between what it gives away, older or smaller base architectures and tooling, and what it monetizes, which is the top of the stack.
The pricing is the other half of the story. At $0.15 and $0.47 per million tokens with steep audio and video discounts, Alibaba is undercutting comparable Western omnimodal endpoints by a wide margin, and doing so during the week of US-China trade talks that include AI on the agenda. For developers building voice agents or video pipelines, the unit economics of a single omnimodal call versus a chain of transcription, vision and language models are now hard to ignore. For Alibaba, keeping the weights closed means that advantage accrues to Alibaba Cloud rather than to whoever runs the model cheapest.
What to Watch
The obvious question is whether independent benchmarks confirm Alibaba's 26-percent claim, and whether tool-use reliability holds across three-hour audio sessions and 113 languages rather than curated evals. Watch for a Qwen-Live Harness repository to actually appear, for any hosted pricing or commercial-license terms on Qwen-Image-2.1, and above all for whether the next Qwen model release follows the Omni-Flash pattern or the open-weights one. If it is the former, the open ecosystem that grew up around Qwen will have to decide where else to go.
"These advancements mean that audio and video are evolving from perceptual inputs to core media for agents to understand and reason about their environment and perform tasks."— Qwen team, Alibaba technical write-up