Qwen 4: what Alibaba announced at Apsara 2026
Qwen 4 was announced at Apsara on 22 September 2026 but is not released: no weights, price or date as of 8 October. What is confirmed, and what to use now.
Qwen 4 was announced on 22 September 2026, but it has not been released. At its Apsara Conference in Hangzhou, Alibaba said the model "is currently in training" and set a roadmap for Qwen 4.5 and Qwen 5 of 5 to 10 trillion parameters. There is no release date, price, context window or downloadable model yet, and as of 8 October 2026 we found no Qwen 4 model card, weights, API name or price. What Alibaba did ship this month is Qwen3.8: an omni-modal Flash model that understands video and audio, and an updated Qwen3.8-Max. It also put out an open-weight preview of the Qwen 4 architecture in August. This page separates the three, and it will be updated when Qwen 4 actually lands.
What Alibaba said about Qwen 4
The official source is Alibaba Cloud's Apsara 2026 press release, issued on 22 September. On Qwen 4 it says two things:
- Qwen 4 "is currently in training".
- The Qwen 4.5 and Qwen 5 series are "projected to scale up to 5 to 10 trillion parameters".
That is all the release confirms. It gives no tier names, no launch window, no benchmarks and no pricing.
The tier names come from the stage. Reports from the conference, including TestingCatalog and Pasquale Pillitteri's write-up, say Qwen's model lead Liu Dayiheng previewed four tiers: Qwen 4 Max, Qwen 4 Plus, Qwen 4 Flash and Qwen 4 27B. He described a new-generation architecture and said the release would come "very soon". Until Alibaba publishes a model card, treat the names as previewed and the timing as unknown.
Two corrections to claims circulating online:
- The 10 trillion figure is not Qwen 4. It is the goal for the generations after it, Qwen 4.5 and Qwen 5.
- "Qwen 4 is out" is wrong. It was wrong on 22 September and we found nothing to change that by 8 October. Anything offering "Qwen 4 access" today is serving a different model.
What you can use now: Qwen3.8
While Qwen 4 trains, the current generation is Qwen 3.8, and three parts of it matter this month.
Qwen3.8-Omni-Flash (released 18 September 2026). A single model that takes text, images, audio and video, with a context window of 1 million tokens, according to TechNode and Qwen's own post. Alibaba pitches it at long-video analysis, meeting summaries and agent work that calls tools. Qwen reports an average gain of more than 25% across 29 evaluations over Qwen3.5-Omni-Plus, and says the API price per hour of audio input fell by more than 98%; TechNode puts input pricing as low as RMB 0.8 per million tokens. For a content team, the interesting part is video understanding: a model that can watch an hour of footage and answer questions about it, or find the moments worth cutting.
Qwen3.8-Max, updated. The press release says Alibaba let the model run its own training experiments for a month (33 cycles), which lifted its Artificial Analysis score from 40 to 45. It is the flagship text model until Qwen 4 arrives.
Qwen3.8-Flash-Next (open weights, August 2026). This is the nearest thing to Qwen 4 you can download. Its Hugging Face model card calls it an "experimental preview of the architecture that will underpin Qwen4". The card lists 125B parameters in total with 6B activated per token, plus a 51B n-gram embedding table, a native context of 262,144 tokens extensible to 1,000,000, text, image and video input, and the qwen-community-1.0 licence. It is a preview, not Qwen 4: Alibaba has not said that the Qwen 4 tiers will match it, and the model card says nothing about a 27B tier.
The rest of the Apsara announcements that matter to creators
The same release lists several media models. None is Qwen 4, but several are closer to a creator's work:
- Qwen-Audio-3.1-TTS-Next, described as next-generation audio generation for "cinematic soundscapes", alongside new Qwen-Audio-3.1 speech recognition, speech and real-time voice models.
- Qwen3.8-LiveTranslate, simultaneous interpretation with latency cut "nearly 20% from 2.8 to 2.3 seconds". That is useful for live streams and for dubbing drafts.
- Qwen-Image 3.1, image generation "later this year", with native transparent-background output. No date or samples yet.
- Happyworld Arena, a benchmarking platform for video generation, embodied AI and 3D reconstruction.
Alibaba also runs a separate "Happy" family of media models outside the Qwen name. HappyShrimp 1.0 is its text-to-song model, in beta since 17 August 2026, which writes lyrics, melody, arrangement and vocals from a prompt (TechNode). HappyHorse is its video model line, and HappyOyster an interactive world model for game-style 3D scenes.
Alibaba's releases around Apsara, dated
The launches blur together in coverage, so here they are in order, with what you can actually use.
| Date | Release | What it is | Status |
|---|---|---|---|
| 17 Aug 2026 | HappyShrimp 1.0 | Text-to-song: lyrics, melody, arrangement and vocals | Beta |
| Aug 2026 | Qwen3.8-Flash-Next | Open-weight preview of the Qwen 4 architecture (125B total, 6B active) | Weights on Hugging Face |
| 18 Sep 2026 | Qwen3.8-Omni-Flash | One model for text, images, audio and video, 1M-token context | Available on Qwen's platform and API |
| 22 Sep 2026 | Qwen3.8-Max (updated) | Flagship text model, Artificial Analysis score 40 → 45 | Available |
| 22 Sep 2026 | Qwen-Audio-3.1 family | TTS-Next, ASR, TTS and Realtime voice models | Announced; availability not stated |
| 22 Sep 2026 | Qwen3.8-LiveTranslate | Simultaneous interpretation, 2.3 s latency | Announced; availability not stated |
| Later in 2026 | Qwen-Image 3.1 | Image generation | Announced for "later this year" |
| Not dated | Qwen 4 (Max, Plus, Flash, 27B) | Next-generation language models | In training |
Why a language model launch matters to a video team
Qwen 4 will not render a frame. Its value to a content team is the same as any strong language model's: writing scripts and captions, planning a campaign, reading a brand's site, and driving an agent that calls the image, video and voice models. The detail worth watching for that job is licensing. Earlier Qwen generations were released with open weights; if Qwen 4 follows, a mid-size tier could run on your own hardware for work you would rather not send to a third party. That is not confirmed for Qwen 4 yet, and price per token will matter just as much once it is published.
Alibaba models you can use in Versely today
Versely does not offer Qwen 4 or Qwen3.8-Omni-Flash. It does carry a large set of Alibaba's media models, priced in credits and shown before you generate:
- Image: Qwen Image 3 text to image and Qwen Image Edit 2511 for changing one thing in a picture without redrawing the rest.
- Video: Wan 3.0 text to video and the Happy Horse line, such as Happy Horse 1.0 text to video.
- Voice: none since 24 September, when Versely retired Qwen TTS in favour of Gemini 3.8 Flash TTS.
When Qwen-Image 3.1 ships with a public API, it is the Alibaba launch most likely to reach the catalog. Language models like Qwen 4 plan and write rather than make pictures, so they belong to the agent side, not the render list.
What to do now
- If you build on Qwen: nothing to migrate yet. Qwen3.8-Max and Qwen3.8-Omni-Flash are current.
- If you make video: try Qwen3.8-Omni-Flash for footage review, and keep your renderer where it is.
- If you are waiting for Qwen 4: watch for a model card and weights. Alibaba has released open-weight versions of past Qwen generations, and a 27B tier would be the one that runs on a single high-end GPU, but nothing about Qwen 4's licence is confirmed. To test the new architecture before then, try Qwen3.8-Flash-Next.
FAQ
Has Qwen 4 been released?
No. On 22 September 2026 Alibaba said Qwen 4 "is currently in training". It has no release date, price or public weights.
What are the Qwen 4 models?
Reports from the Apsara Conference say four tiers were previewed on stage: Qwen 4 Max, Qwen 4 Plus, Qwen 4 Flash and Qwen 4 27B. Alibaba's press release does not list them.
When is the Qwen 4 release date?
Not announced. Qwen's model lead said "very soon" on stage, with no date.
Is Qwen 4 10 trillion parameters?
No confirmed size. The 5 to 10 trillion parameter figure is Alibaba's target for Qwen 4.5 and Qwen 5, the generations after Qwen 4.
What is the newest Qwen model I can use?
Qwen3.8-Omni-Flash, released on 18 September 2026, and the updated Qwen3.8-Max. If you want the Qwen 4 architecture itself, the open-weight Qwen3.8-Flash-Next preview from August 2026 is the closest thing to it.
Is there a Qwen 4 preview I can download?
Yes, in architecture only. Qwen3.8-Flash-Next is on Hugging Face and its model card calls it an experimental preview of the architecture that will underpin Qwen4. The Qwen 4 models themselves, including the 27B tier, have no model card or weights yet.
Can I use Qwen models in Versely?
Yes, Alibaba's image, video and voice models: Qwen Image 3, Qwen Image Edit 2511, Wan 3.0 and Happy Horse. Qwen TTS was retired on 24 September. Qwen 4 and Qwen3.8 are not in Versely.
Versely AI Weekly
New AI models, every Thursday.
The image, video, audio and LLM models released that week, what each is good for, and what's new on Versely.