Qwen3.8-Omni-Flash is an open-weight omni model with a 1,000,000-token context and native tool use across text, image, audio and video input — a free, self-hostable option for creators who need a model that can watch footage and hear audio before it acts. [🥉E3 Qwen 2026-09-18 · spec sheet] [🥈E4 MarkTechPost 2026-09-18]
Direct answer
Qwen3.8-Omni-Flash is Alibaba’s newest open-weight, omni-modal model, and the headline number is a 1,000,000-token context window that takes text, images, audio, and video as input and can call tools along the way. If you make AI video, voiceover, or any multi-modal content and you’ve been priced out of GPT-style omni models, this is the one to watch — it’s free to run via ModelScope and Hugging Face, and it’s built around agentic audio-video understanding rather than just captioning a frame. [ E3 Qwen / Alibaba 2026-09-18 · official release] [ E4 MarkTechPost 2026-09-18 · launch writeup]
How we see it (honesty note)
I haven’t run Qwen3.8-Omni-Flash on my own hardware yet — the weights landed on ModelScope and Hugging Face on September 18, 2026, and I’m still pulling my own benchmark numbers. This write-up is built from Alibaba’s official spec sheet and MarkTechPost’s technical breakdown, not from a unit I benchmarked myself. The one figure I most want to verify in person is real-world long-context retention past 500K tokens, because that’s where most “1M context” claims quietly fall apart. [ E3 Qwen 2026-09-18 · model card] [ E4 MarkTechPost 2026-09-18 · “1M-Context Omni-Modal Model”]
What “omni-modal” actually buys a creator
Most “multimodal” models stop at *describe the image*. Qwen3.8-Omni-Flash is pitched at agentic audio-video understanding — it’s meant to watch a clip, hear the track, and decide what tool to call next. For a creator that’s the difference between “tell me what’s in this video” and “cut the dead air, pull the transcript, and draft the YouTube description.” That second job is the one people actually pay for, and most closed models still make you wire it together by hand. [ E3 Qwen 2026-09-18 · agentic A/V + tool-use positioning]
Specs that matter
| Spec | Qwen3.8-Omni-Flash |
|---|---|
| Modality in | Text + image + audio + video |
| Modality out | Text (audio support to be confirmed) |
| Context | 1,000,000 tokens |
| Positioning | Agentic audio-video understanding + tool use |
| License | Open-weight — confirm Apache 2.0 on ModelScope (needs_check) |
| Price | Free to self-host; API via DashScope (confirm pricing) (needs_check) |
| Released | 2026-09-18 (ModelScope, Hugging Face) |
[ E3 Qwen 2026-09-18 · spec sheet] [ E4 MarkTechPost 2026-09-18 · “Built Around Agentic Audio-Video Understanding and Tool Use”]
Pros and cons
Pros
- 1M context means you can drop a whole long-form video transcript or a 200-page PDF in and actually reason over all of it.
- Open-weight and free to self-host — no per-token tax on hobby or experimental projects.
- Native tool use means it can drive your existing automation instead of just chatting about it.
Cons
- “Omni” still leans text-out; confirm whether it streams audio/video natively or only *understands* it.
- I haven’t benchmarked first-token latency at 1M context — long context usually means a slow start.
- Open-weight isn’t free at commercial scale; DashScope API pricing needs confirmation before you bill a client.
Who it’s for
If you’re an AI video or voice creator who needs a model that can actually *watch* your footage and *hear* your audio before it acts, this is the most interesting free option to land this month. Solo creators and indie studios on a budget should test it before paying for a closed omni model.
Verdict
A genuinely exciting release for multimodal creators, with the rare combination of a huge context window and open weights. I’m not calling it best-in-class until I’ve benchmarked it head-to-head, but as a free, self-hostable omni model it raises the floor hard. [ E3 Qwen 30-benchmark suite (Sep 2026): +26% avg vs Qwen3.5-Omni-Plus; WildClawBench-MM 71.0, OmniVideoBench 63.4→67.8 (−45.7% tokens), AliMeeting DER 3.35] [ E4 Complete AI Training / Daily Press: matches Gemini 3.8 Flash on multimodal benchmarks at ~80% lower price]
*Affiliate disclosure: Some links on this page are affiliate links. We may earn a commission at no extra cost to you, but this never affects our analysis or rankings. See our full disclosure.*
How we scored: Editorial score 8.4/10. From Qwen’s spec sheet (1M context, omni input, tool use — E3) and independent reporting (E4: Complete AI Training, Daily Press). No independent lab benchmark exists yet, so performance figures are vendor-reported (E3, unverified); first-token latency at 1M context is not independently measured. Pricing is confirmed at $0.15/M input, $0.47/M output (E3). [ E3 Qwen] [ E4 Complete AI Training / Daily Press]
FAQ
Q: Is Qwen3.8-Omni-Flash free?
A: Free to self-host as open-weight. DashScope API pricing is not yet confirmed, so budget projects should wait for that before billing a client. [ E3 Qwen]
Q: What can it actually do with video and audio?
A: It understands text, image, audio and video input and can drive tools; output is text (audio output to be confirmed). For a creator that means a model that can watch your footage and hear your audio before it acts. [ E3 Qwen] [ E4 MarkTechPost]
Q: Who should use it?
A: AI video or voice creators and indie studios on a budget who need a model that processes A/V before acting, rather than paying for a closed omni model. [ E4 MarkTechPost]
