The release calendar moved faster than the headlines
Late one week in September, Alibaba's Qwen team put three models into the world within three days. One of them you can download and run yourself. The other two you can only rent. That split is the real story.
Between 18 and 20 September 2026, Qwen published Qwen-Image-2.1 (open weights), Qwen3.8-Omni-Flash (API), and Qwen3.8-LiveTranslate (API). For a company that spent years keeping its best models behind a paywall, the cadence and the open/closed mix say something about how Chinese frontier labs now compete.
What actually shipped
Qwen-Image-2.1 — the one you can own. Released 20 September 2026 and open-sourced on GitHub, Hugging Face and ModelScope. Its visual-generation component is small: 32 single-stream DiT layers and 7 billion parameters. It unifies text-to-image generation and image editing in one model, adds native transparent-image support (generate and edit RGBA layers), and accepts up to 10 reference images for a single composition. A mixed-granularity attention design with KV-cache reuse is what keeps it cheap to run.
Qwen3.8-Omni-Flash — the one you rent. Released 18 September 2026 as a hosted API, not open weights. It takes text, image, audio and video and returns text, behind a 1-million-token context window (max input about 991k, max output 131k). It understands audio in 113 languages and dialects, handles video up to roughly 2 hours / 2 GB, and ships with function calling, web search and context caching. Alibaba says it is built around "agentic" workflows — find the relevant moment in a long video, then call tools to edit or summarise it.
Qwen3.8-LiveTranslate — the one for your ear. Also released 18 September 2026. Alibaba says its average latency dropped from 2.8 seconds to 2.3 seconds (LAAL) versus the previous generation. Alibaba Cloud's documentation describes it as understanding 60 input languages and speaking 29 output languages through a realtime WebSocket interface that accepts audio and images.
The fine print that matters
Here is the part easy to miss in a launch blog: only Qwen-Image-2.1 ships weights. Both Omni-Flash and LiveTranslate are API-only. Alibaba open-sourced companion tooling — Qwen-MM-Plugins under an Apache-2.0 licence — so developers can bolt multimodal abilities onto existing agent harnesses, but the models themselves stay closed.
That is a deliberate split, not a slip. Image generation is now a crowded, fast-commoditising space where giving weights away builds adoption and an ecosystem. Long-context multimodal understanding, by contrast, is where Alibaba is trying to charge: the international API price for Omni-Flash is listed at US$0.15 per million input tokens and US$0.47 per million output tokens, served from six regions including Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia.
Why the cost numbers are the headline
For the omni model, Alibaba's own claims are about price, not just capability. It says the hourly API cost of audio input fell more than 98% and combined audio-plus-video input fell more than 93% versus the prior generation. If those vendor figures hold up in production, long video and meeting audio stop being expensive to analyse and start being a default input — the difference between a feature you demo and a workflow you run.
The catch, again: those are Alibaba's benchmark and cost claims, and independent evaluations for the new endpoints were not yet published at launch. The honest read is that the price drop is plausible given the architecture, but unproven by outside tests.
Honest limitations
- Only one of the three models is open source. Qwen-Image-2.1 ships weights; Qwen3.8-Omni-Flash and Qwen3.8-LiveTranslate are API-only, despite "release" framing that can read as open.
- Performance and cost claims are vendor-reported. The 98%/93% cost reductions and the 2.3-second latency are Alibaba's numbers; we found no mature independent benchmark at publication.
- Some specifications come from secondary coverage. The 113-language audio support and 1M context are confirmed in Alibaba Cloud documentation and Sina reporting; the 60/29 language split for LiveTranslate is from Alibaba Cloud docs via a secondary summary and should be confirmed in-console.
- Pricing is the international rate. Domestic China pricing differs and was not quoted here; token costs vary by region, so a production budget needs the console figure for your deployment region.
- No renminbi (RMB) amounts appear in this piece because the cited API pricing is published in US dollars; we did not convert or invent domestic RMB figures.
What readers can do now
- If you ship image-heavy product or content work, download Qwen-Image-2.1 and benchmark it against your current tool — the 7B component and transparency support are genuinely useful for design, e-commerce and storyboarding, and you can run it privately.
- If you build video or meeting agents, test Omni-Flash on one long file you currently have to chunk or transcribe separately, and measure real token cost in your region before trusting the headline price drop.
- Don't conflate "Alibaba open-sourced a model" with "Alibaba gave away its AI." This month it gave away the image model and rented the rest — read the licence, not the headline.