Quick answer
The Qwen 4 preview is an early-access version of Alibaba’s next model family, which includes Qwen 4 Max, Qwen 4 Plus, Qwen 4 Flash and Qwen 4 27B. WorldofAI got preview access through Qwen Code and showed detailed coding demos, including a Minecraft-style game, a Roman Colosseum scene and an insulin bioreactor simulation. Early tests suggest strong web development skills, better results at higher reasoning levels and about half the tokens of DeepSeek V4.1 Flash on one task. A deleted partner image hinted at a 1 million token context window for Qwen 4 Max. Alibaba has not confirmed the release date, sizes behind each demo or pricing.
Key takeaways
- Four Qwen 4 models were announced at Alibaba’s Apsara conference: Max, Plus, Flash and 27B, with a Qwen Coder model also hinted.
- An Alibaba Cloud CTO reportedly said Qwen 4 was coming “next week” on 29 October, pointing to a launch window of 2 to 8 November.
- A deleted partner image listed a 1 million token context window, image input, tool calling and a free-token launch promotion for Qwen 4 Max.
- Demos show strong web development and agentic coding, with results the presenter compares to Claude Sonnet and Opus.
- Qwen 4 reportedly improves at x-high and max reasoning, but takes longer.
- Arena checkpoints such as Kerria, Kalmia and Mirra are believed to be Qwen 4 variants.
- Qwen 4 is expected to be open-weight, and a rumour that Alibaba is dropping smaller open models is disputed.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- Low, High or Max: How to Choose an AI Reasoning Effort Level — how to pick between low, high and max thinking without wasting time or tokens.
- How to Compare AI Models on Token Cost, Not Just Output Quality — a simple way to test claims like “half the tokens of DeepSeek” on your own tasks.
In this guide
- Watch: the source video
- What is the Qwen 4 preview?
- When is Qwen 4 coming out?
- What did the Qwen 4 Max leak reveal?
- What did the early Qwen 4 demos show?
- How does Qwen 4 compare with GPT-6.1 Terra, DeepSeek and Claude?
- Does Qwen 4 really reason better and use fewer tokens?
- Will Qwen 4 be open-weight?
- How can you try Qwen 4 now?
- What does the video not confirm?
- How should you prepare for Qwen 4?
- Who is Qwen 4 right for?
- Frequently asked questions
- The bottom line
- Related posts
Alibaba’s Qwen models have become the go-to open models for many builders, and Qwen 4 has been teased for weeks. Now early testers have it, and the first demos are out.
This article is a first look at the Qwen 4 preview: what the family includes, what leaked about Qwen 4 Max, what the early coding demos show and how it compares with GPT-6.1 Terra, DeepSeek V4.1 Flash and Claude Opus 5.5. It sticks to what the video shows and flags what is still unconfirmed. For earlier Qwen 4 rumours inside wider news round-ups, see our 5 November AI news round-up.
Watch: the source video
What this video covers: WorldofAI’s first look at the Qwen 4 preview covers the Qwen 4 Max, Plus, Flash and 27B line-up, the expected release window, a deleted Qwen 4 Max specification leak, Alibaba’s teaser poster, and hands-on demos built in Qwen Code and other tools, with comparisons against GPT-6.1 Terra, DeepSeek V4.1 Flash, GPT-6 Astra and Claude. It ends with the rumour about Alibaba’s smaller open-weight models. The video is sponsored by Krisp.
Video: WorldofAI (@intheworldofai). Sponsored by Krisp. This summary and the opinions in it are our own. Pre-release results, leaked specifications and release dates are unconfirmed until Alibaba announces them.
What is the Qwen 4 preview?
It is early access to Alibaba’s Qwen 4 model family, given to partners and beta testers before a public launch. Alibaba introduced the series at its Apsara conference about a month before the video.
| Model | What we know |
|---|---|
| Qwen 4 Max | The largest model. Leaked details suggest a 1 million token context window. |
| Qwen 4 Plus | A mid-sized model. WorldofAI thinks the Qoduck stealth model may be this one. |
| Qwen 4 Flash | A faster, lighter option. |
| Qwen 4 27B | A smaller model with 27 billion parameters, the one the presenter is most excited about. |
| Qwen Coder (hinted) | Traces suggest a dedicated coding model could return later. |
One important caveat: the preview did not say which size answered each request. In a follow-up video, the presenter explained that the preview API routed requests across the Qwen 4 models, so some demos may come from the 27B model and others from a larger one. We cover that in our Claude Silicon and Qoduck round-up.
When is Qwen 4 coming out?
Soon, but there is no official date. According to the video, Alibaba Cloud’s CTO said at the Apsara event in Amsterdam on 29 October that Qwen 4 was coming “next week”, which suggested 2 to 8 November.
The presenter also says early testers are expected to get a few days with the model before a wider rollout, possibly the following week. Alibaba has posted teasers too, including one called “The becoming begins” and a “Qwhen? Qsoon!” post.
🔎 Unconfirmed / rumour: The release window comes from a reported comment at a conference and from what the presenter has heard privately. Until Alibaba publishes a launch post, treat any date as a guess.
What did the Qwen 4 Max leak reveal?
A partner reportedly posted, then deleted, a promotional image listing Qwen 4 Max specifications. According to the video, it showed:
- A 1 million token context window, roughly enough for several long books in one prompt.
- Image input, so you can send screenshots or photos.
- Tool calling, which lets the model use external tools and APIs.
- A launch promotion with free input and output tokens.
Alibaba’s own teaser poster adds a playful clue. Its Chinese text quotes an ancient philosophical classic about Kun, a giant fish in the northern sea so large nobody knows its size. The presenter reads this as a hint that Qwen 4 is big. It could just as easily be clever marketing.
What did the early Qwen 4 demos show?
The demos show a model that is strong at building detailed web apps, games and simulations from a single prompt. Here is what the video shows, as described by the presenter:
| Demo | What it showed |
|---|---|
| Minecraft-style game (Qwen Code) | Shaders, day and night, inventory, water physics, caves, mobs, sound and a creative mode. One small bug: the player could not leave the water. |
| Roman Colosseum | A detailed 3D arena with crowds and flickering lights, at high (not max) thinking, using about 69,000 tokens. |
| 2D RPG village | Kerria checkpoint vs GPT-6.1 Terra, both at max. The presenter prefers Qwen’s layout and feel. |
| Insulin bioreactor (Hermes harness) | Wrote and tested a Python growth model, then built an interactive simulator with oxygen, temperature, pH and feed controls. |
| Call of Duty-style shooter | Basic shooting and enemy behaviour at a lower reasoning level. |
| Voxel scene, boat scene and aquarium | Day-night cycles, water effects and lighting in Three.js. The presenter says the aquarium beat his GPT-6 Astra attempt. |
| SVG animation | A giraffe riding a bike with a rhino on top, compared to Sonnet 5.5 quality. |
The bioreactor simulation reported 18.4 grams of proinsulin from a 36.6-hour batch. The presenter is clear that this is a simulation, not a validated production figure.
⚠️ Warning: Impressive one-prompt demos are a useful signal, but they are not the same as reliable everyday work. Several demos had small bugs, and none of them is an independent benchmark.
How does Qwen 4 compare with GPT-6.1 Terra, DeepSeek and Claude?
Early tests put Qwen 4 close to top closed models on creative coding, but these are hand-picked comparisons, not benchmarks.
- Vs GPT-6.1 Terra: on a 2D RPG village, the presenter preferred Qwen’s composition, while saying Terra is not bad.
- Vs DeepSeek V4.1 Flash: better web development output and roughly half the tokens. Read our DeepSeek V4.1 Flash test for background.
- Vs Claude Opus 5.5: in a dungeon crawler test shown in the follow-up video, the presenter called the difference “very marginal”.
- Vs Qwen 3.8: a clear step up in design taste and complexity, according to the presenter.
Planning to test Qwen 4 yourself? Start with how to choose an AI reasoning effort level, then use our token cost comparison method so you judge it on your own work.
Does Qwen 4 really reason better and use fewer tokens?
Early signs say yes to both, with one trade-off: it is slower at higher effort. The presenter says that moving Qwen 4 from high to x-high or max gives noticeably better results, which was not always the case with earlier Qwen models. It also takes longer, and he says it reasons for longer than GPT or Opus models.
On tokens, the DeepSeek comparison and the boat scene both suggest Qwen 4 uses fewer tokens for similar or better results. Even so, the dungeon crawler at max used about 200,000 tokens, so heavy tasks can still be expensive once normal pricing applies.
Two guides help you test this properly: Low, High or Max: How to Choose an AI Reasoning Effort Level and How to Compare AI Models on Token Cost, Not Just Output Quality.
Will Qwen 4 be open-weight?
It is expected to be, based on Alibaba’s Apsara conference slides, which the presenter says described both Qwen 4 and Qwen 5 as open-weight. Open-weight means you can download the model files and run them on your own hardware.
A rumour from Chinese online communities claimed Alibaba might stop investing in smaller open models. The presenter believes it is false, saying the person who posted it had no link to the Qwen team. If you want to run a model like Qwen 4 27B locally, our guide on how much memory you need to run an open-weight AI model is a good place to start.
How can you try Qwen 4 now?
You cannot get the official early-access preview, but there are two routes to try related models.
- Arena battle mode: recent checkpoints such as Kalmia and Mirra, believed to be Qwen models, may appear in anonymous battles. Our guide on testing unreleased AI models in Arena battle mode explains how.
- Qoduck on Qoder: a free stealth model that testers think is probably a Qwen 4 model. Read how to use a free stealth AI model without risking your data first.
What does the video not confirm?
🔎 Unconfirmed / rumour: The video does not confirm the release date, which model size produced each demo, final pricing, whether the free-token promotion is real, official benchmark scores, or that the Arena checkpoints and Qoduck are Qwen 4. The Qwen 4 Max specifications come from an image that was deleted. The open-weight plans come from conference slides as described by the presenter.
How should you prepare for Qwen 4?
- Write a short test pack of three real tasks you do often.
- Run it now on your current model, noting quality, time and tokens, so you have a baseline.
- Decide your effort levels using our reasoning effort guide.
- Set up a token cost sheet using our comparison method.
- Test Qwen 4 on launch day with the same pack and settings.
- Check the data terms for whichever host you use: Alibaba Cloud, a third-party API or your own machine.
- Switch only if it wins on cost per usable result for a specific task.
Who is Qwen 4 right for?
Qwen 4 suits you if you build websites, apps or games with AI, want a strong open-weight option you can host yourself, or are watching API costs closely.
It matters less if you are happy with your current assistant for writing and email, or you do not want to manage different providers. You can wait for independent reviews after launch.
Ready to be ready for launch day? Set your defaults with our reasoning effort guide, and build a fair test with our token cost comparison.
Frequently asked questions
What models are in the Qwen 4 family?
Alibaba announced Qwen 4 Max, Qwen 4 Plus, Qwen 4 Flash and Qwen 4 27B. A dedicated Qwen Coder model has also been hinted.
When will Qwen 4 be released?
There is no official date. An Alibaba Cloud CTO reportedly said “next week” on 29 October, and early testers expect a wider rollout soon after their preview period.
Does Qwen 4 Max have a 1 million token context window?
A deleted partner image said so, along with image input and tool calling. Alibaba has not confirmed it.
Is Qwen 4 better than DeepSeek V4.1 Flash?
In WorldofAI’s early tests, the Qwen 4 preview produced better web development output with about half the tokens. These are pre-release tests, so check again after launch.
Will Qwen 4 be open source?
It is expected to be open-weight, based on Alibaba’s conference slides as described in the video. A rumour that Alibaba is dropping smaller open models is disputed.
Should I use higher reasoning effort with Qwen 4?
Early testers say Qwen 4 improves at x-high and max, but it is slower. See how to choose an AI reasoning effort level.
Who sponsored the WorldofAI video?
The video is sponsored by Krisp, an AI meeting note taker with noise cancellation.
The bottom line
The Qwen 4 preview looks like a big step up for Alibaba’s open models, with strong coding demos, better results at higher reasoning levels and promising token efficiency. But it is still a preview: the release date, which size made each demo, pricing and the Qwen 4 Max specifications are all unconfirmed. Build a short test pack and a token cost sheet now, then judge Qwen 4 on your own work when it launches.
Related posts
- DeepSeek V4.1 Flash 1017 Tested: How a Cheap Open Model Stacks Up
- How Much Memory Do You Need to Run an Open-Weight AI Model?
- How to Test Unreleased AI Models in Arena Battle Mode
