Qwen 3.8 Omni Flash now powers Vega's tools
Alibaba released Qwen 3.8 Omni Flash on September 18, 2026. It is a natively omni model: text, images, audio, and video go in, text comes out. The context window is 1 million tokens, and thinking mode is on by default with an adjustable effort level.
We upgraded the model behind Vega's backend work to Omni Flash. It runs the web search groups, file and document analysis, transcription summaries, and tool calls on vega.chat and ru.vega.chat. There is nothing to select in the model picker: these features already use the new model.
What omni adds
The previous Qwen 3.8 Flash read text and images. Omni Flash also understands audio and video natively, including two- and four-channel spatial audio, where it can estimate the direction and distance of a sound source. For file uploads this means recordings and clips are handled by a model that hears and watches the content itself.
Alibaba frames the release around agentic execution: the model plans a task, calls tools, and returns a finished result. Its demo scenarios include video editing, clip assembly, film commentary, and summaries of long tutorials. That profile fits how Vega uses the model, since it drives search groups and multi-step tool calls behind the chat interface.
Alongside the model, Alibaba published the open Qwen-MM-Plugins for agent frameworks and Qwen-Live-Harness, an open-source environment for realtime apps over WebSocket and WebRTC.
Benchmarks
On WildClawBench-MM, a benchmark for multimodal tool use, Omni Flash scores 71.0 against 58.9 for Gemini 3.8 Flash and 34.5 for the previous Qwen3.5-Omni-Plus. On OmniVideoBench it reaches 63.4 in static mode and 67.8 in Qwen Code mode, while using 45.7% fewer tokens than its predecessor. Alibaba reports an average gain of more than 25% across about 30 benchmarks.
Pricing
Text pricing matches the previous Flash: $0.15 per million input tokens, $0.47 per million output tokens, and $0.016 per million tokens on cache hits. Audio input is more than 98% cheaper than in the previous omni generation, and combined audio-video input is more than 93% cheaper.
Specs
| Parameter | Value |
|---|---|
| Input | text, images, audio (2ch/4ch spatial), video |
| Output | text |
| Context | 1M tokens, 983K in thinking mode |
| Max output | 131K tokens |
| Thinking | on by default, adjustable effort |
| API | DashScope and OpenAI-compatible, 2M TPM / 30K RPM |
Where it works on Vega
Upload an audio file or a video clip, ask a question that triggers web search, or send a document for analysis. The answers for these tasks now come from Qwen 3.8 Omni Flash.