Qwen3.8-Omni-Flash: 1M Context & API Pricing (2026)

Qwen3.8-Omni-Flash is an omni-modal model that Alibaba's Qwen team announced on September 18, 2026. It accepts text, image, audio and video as input, offers a 1-million-token total context, and starts at $0.15 per million input tokens. Its weights are closed and proprietary, so it cannot be run locally.

Unlike its predecessor Qwen3.5-Omni-Plus, which could also generate audio output, Qwen3.8-Omni-Flash's output is text only. Anyone who needs a spoken response has to build their own multi-stage pipeline with an external text-to-speech engine bolted on afterward.

It is hosted on the Qwen AI platform (Tongyi/Qianwen) and Alibaba Cloud Model Studio (Bailian), both accessed via API. Below is a rundown of its specs, pricing, vendor-reported benchmarks, and how it compares with existing models.

- Long-video analysis (summarizing meeting recordings, surveillance footage, lecture videos)
- Meeting summarization (generating minutes and key points from audio)
- Video research (cross-referencing information across multiple videos)
- Agentic tasks involving multimodal tool use

All figures are as published by Alibaba; actual billed rates can vary by region and contract. For a sense of pricing across the Qwen line, see our Qwen3.7-Flash API pricing guide.

All of the above figures come from Alibaba; no independent third-party evaluation exists yet. Real-world performance should be verified against your own workloads. Alibaba also released Qwen-MM-Plugins and the Qwen-Live Harness alongside the model.

- Issue an API key on the Qwen AI platform (Tongyi/Qianwen) or Alibaba Cloud Model Studio (Bailian)
- Send a request specifying Qwen3.8-Omni-Flash as the model (check the official docs for the current exact model ID and parameter names)
- When sending audio or video, confirm file-size limits and the one-hour continuous-media cap in advance
- Since output is text only, design a multi-stage pipeline with an external TTS engine downstream if a spoken response is needed
- For long-context workloads, keep shared system prompts or context as a fixed segment so it qualifies for cached-input pricing

- Choose it: you need to process large multimodal input — long video, meeting audio — cheaply
- Choose it: you want to use the 1M-token context to reason across multiple videos or documents
- Choose it: text-only output is sufficient for your use case (summarization, analysis, search)
- Don't choose it: you need direct spoken responses (no audio output like the predecessor, so a multi-stage pipeline is mandatory)
- Don't choose it: you need to self-host the weights or require local deployment (consider open-weight alternatives like Qwen3.8 27B or Qwen3.8-Flash-Next)
- Don't choose it: your procurement requirements mandate independently verified benchmarks (only vendor-reported figures exist so far)

Qwen3.8-Omni-Flash is a strong candidate for multimodal workloads involving long audio and video, backed by its 1M-token context and low pricing. But its closed weights ruling out local deployment, and its text-only output, are constraints worth weighing carefully against your requirements before adopting it.