Qwen-4 Preview: Official Teaser, Four Tiers, 10T Parameter Roadmap, and the Open-Weights Question
Qwen-4 Preview: Official Teaser, Four Tiers, 10T Parameter Roadmap, and the Open-Weights Question
Japanese tech media has been covering Qwen-4's movements in a burst. The trigger was Alibaba's Apsara 2026 conference in September, where the company unveiled the Qwen-4 family structure, a roadmap pointing at 10-trillion-parameter scale, and a new AI chip called Zhenwu V900. In November 2026, an official teaser appeared, and reports say integration testing with Apple began on November 4.
This article separates what is officially confirmed from what is third-party benchmark coverage, then maps it to two audiences: developers who want to run GGUF locally, and developers who want to call a cloud API. The goal is to tell you what you can actually do today, without waiting for the full picture.
The 30-second version
Qwen-4 was announced at Apsara 2026 with a four-tier lineup and a 10T parameter roadmap, and an official teaser followed in November. Qwen3.8-Flash-Next already previews the Qwen4 architecture and is available now, including through cloud APIs. Qwen-4 itself has not confirmed open weights yet.
What was officially announced at Apsara 2026
Alibaba used its September 2026 Apsara conference to present the Qwen-4 family structure and the direction of its parameter scale. Multiple outlets describe a four-tier lineup (OrcaRouter). Reuters reported that the same event unveiled a new AI chip and a larger model, with shares rising 5% (Reuters).
On parameter scale, several outlets report a 10-trillion (10T) roadmap:
- TechRepublic: "Alibaba Unveils Zhenwu V900 — and Plans Qwen Models With Up to 10 Trillion Parameters" (TechRepublic)
- Vietnam Investment Review: "Alibaba targets 10 trillion parameters with next-generation Qwen 4 model" (VIR)
- Pandaily: "Alibaba Puts Qwen4 Family Into Training; Roadmap Points to 5–10T Qwen4.5 and Qwen5" (Pandaily)
- Tech in Asia: "Alibaba details AI roadmap as Qwen 4 enters training" (Tech in Asia)
Official announcement vs media interpretation
All of these trace back to Alibaba's official conference. But phrasing like "four tiers" and "10T roadmap" is each outlet's interpretation. For exact specs, follow Qwen's official channels (qwen.ai and the Qwen organization on HuggingFace).
Qwen3.8-Flash-Next: the Qwen4 architecture, already available
Before the full Qwen-4 picture settles, Alibaba has released a model that previews the Qwen4 architecture. That model is Qwen3.8-Flash-Next.
Several outlets describe it as an early look at what Qwen 4 will offer:
- The New Stack: "Alibaba just released Qwen3.8-Flash: an early preview of the architecture in Qwen4" (The New Stack)
- MarkTechPost: "Qwen3.8-Flash-Next: A 125B Multimodal MoE With 6B Active Parameters Previewing the Qwen4 Architecture" (MarkTechPost)
- Unite.AI: "Qwen3.8-Flash-Next Previews Qwen4 Architecture With 6B Active Parameters" (Unite.AI)
- TechNode: "Alibaba's Qwen to open-source Qwen3.8-Flash-Next, previewing Qwen4 architecture" (TechNode)
Japanese outlet Ledge.ai reports that Qwen3.8-Flash-Next puts the next-generation Qwen4 architecture into service early, with training cost around one ninth (Ledge.ai).
What this means for developers
You do not have to wait for the full Qwen-4 reveal. Qwen3.8-Flash-Next is already open and downloadable from HuggingFace. Run it as GGUF locally and you can try Qwen4-generation behavior early, before the flagship itself ships.
Third-party benchmarks (treat as coverage)
Some Japanese media report that pre-release Qwen-4 beats both Claude Opus and Sonnet 5.5 on third-party benchmarks (shiritomo, via Google News Japan).
This is third-party benchmark coverage. This article does not verify or endorse it. Benchmarks shift with conditions, and they are not the same as official evaluation. Test on your own tasks before making decisions.
Separate coverage from verification
"Beats X" headlines are often written for clicks, and frequently omit the test conditions. Before judging Qwen-4's performance, try an existing open model like Qwen3.8-Flash-Next on your own workload. That is the reliable path.
The open-weights question
The biggest draw of the Qwen series has always been open weights. For Qwen-4 itself, however, open weights are not confirmed.
- tech-insider.org: "Qwen4: 4 Variants Named, 0 Open Weights Confirmed" (tech-insider.org)
- Asia Times: "What Alibaba left blank about Qwen4" (Asia Times)
- shattered.io: "Qwen 4 Launch: Alibaba Open-Source Promise in Doubt" (shattered.io)
- Startup Fortune: "Qwen4 Nears Launch as Alibaba's Open Source Promise Faces Real Doubt" (Startup Fortune)
Impact on local GGUF users
Until Qwen-4 open weights are confirmed, the local option for the latest generation is Qwen3.8-Flash-Next. Community GGUF versions have started appearing. If Qwen-4 weights are officially released, the same GGUF workflow will likely apply, but there is no confirmed schedule yet.
Apple integration and the ecosystem
Qwen is expanding beyond the model itself. TechNode reports that Apple confirmed Qwen integration for its China AI rollout, with beta testing starting November 4 (TechNode). BusinessToday Malaysia also reports "Alibaba To Integrate Qwen AI Into Apple iOS From Nov 4" (BusinessToday Malaysia).
This is China-market specific, but it signals that the Qwen family is being embedded into large platforms. For global developers, it is one more input when choosing a model family.
What you can do today: cloud API and local
While the full Qwen-4 picture develops, there are two things you can do now.
Use the cloud API: EXAIV provides an OpenAI-compatible endpoint. Point the base_url at https://exaiv.com/api/ai/v1 and set the model to a platform model ID. Check the live catalog at /models for the current Qwen IDs.
curl https://exaiv.com/api/ai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "qwen3.8-flash",
"messages": [{"role": "user", "content": "Introduce yourself in one sentence"}]
}'
Run it locally: download the Qwen3.8-Flash-Next GGUF from HuggingFace and start it with Ollama or llama.cpp. See the existing guides for command examples.
For pricing, always check /pricing for live rates. This article deliberately contains no price numbers.
When Qwen-4 arrives
Once Qwen-4 is added to /models, switching is a one-line change: update the model field to the new ID. On the local side, wait for official GGUF releases and follow the same steps. The principle does not change: the live /models catalog is the source of truth for model IDs.
Preparing your codebase for Qwen-4
You do not need to rewrite anything for Qwen-4, but a little preparation makes the switch painless when it lands.
Keep model IDs in configuration, not in code. If your application reads the model name from an environment variable or config file, switching to Qwen-4 is a one-line change. If the model name is hardcoded in five places, it is a refactor.
Run a regression set now. Take five to ten real production prompts and record the outputs of your current model. When Qwen-4 becomes available, replay the same prompts and compare. This is the only reliable way to judge whether the new generation is actually better for your workload, rather than relying on headline benchmarks.
Design for model routing. A tiered setup, where easy steps go to a lightweight model and hard steps go to a flagship, does not depend on which specific model wins the next benchmark. When Qwen-4 ships, you can slot it into the flagship tier and keep the rest of the pipeline unchanged.
Watch the open-weights signal. The local path depends entirely on whether weights are released. Until there is an official announcement, treat Qwen3.8-Flash-Next as the latest locally runnable generation, and keep your GGUF workflow generic so a new model can be dropped in without rework.
Why this matters
Teams that prepare a regression set and a configurable model ID can evaluate Qwen-4 within hours of release. Teams that hardcode model names and rely on benchmark headlines usually take weeks to reach the same confidence. The preparation is cheap; the delay is not.
FAQ
Is Qwen-4 available yet?
Will Qwen-4 ship open weights?
Is it true that Qwen-4 beats Claude on third-party benchmarks?
Are there GGUF versions of Qwen3.8-Flash-Next?
Where do I check the cloud Qwen model IDs?
Where do I see pricing?
Related reading
- Qwen3.8-Flash-Next Local GGUF Deployment
- Qwen Model Selection Guide
- Qwen GGUF Local Deployment Guide
- Qwen Architecture: From Transformer to MoE
- Qwen API Troubleshooting
- Model catalog