World of AI's Qwen 4 preview shows coding, 3D and simulation outputs attributed to early Qwen 4 access. They are exciting examples, but the model behind the anonymous tests is not established by the clips. Alibaba has publicly said Qwen 4 is in training; a downloadable Qwen 4 release is a different milestone. That distinction determines what you can test or install today.
Watch the preview for ideas about what to test, not proof that Qwen 4 beats every frontier model. As of 7 November 2026, we found no official Qwen 4 weights, public API identifier, price or model card. If you want a local model now, Alibaba's Qwen3.8-27B is available and has a documented Ollama install route. It is a useful baseline, not the previewed Qwen 4 checkpoint.
Watch the Qwen 4 Preview
The episode includes a sponsored Krisp segment at about 4:00. The model comparisons and previews are the creator's demonstrations, not tests reproduced by this article.
Source Note: What Was Checked
We used the supplied video transcript for the creator's claims and chapter times. On 7 November 2026, we checked Alibaba's Apsara announcement, the official Qwen model publisher, the Qwen3.8-27B model card, Ollama's model listing and Arena's explanation of anonymous testing. We did not receive beta access, identify the anonymous checkpoint, reproduce the coding comparisons or validate the bioreactor's scientific output. Several X links in the supplied description were truncated after /status/; below we link the supplied accounts, not guessed posts.
Confirmed Release or Preview Claim?
| Item | What we can verify | How to read it |
|---|---|---|
| Qwen 4 family | Alibaba says it is in training. | An announced roadmap, not a public model release. The video's Max, Plus, Flash and 27B labels need individual official model cards before comparing specs. |
| Kerria, Kalmia, Kentia, Mura | Arena tests anonymous prerelease models; the video associates these names with Qwen. | A codename or impressive output does not prove the checkpoint belongs to Alibaba, or which Qwen size it is. |
| One-million-token context, tools, free tokens | A deleted partner image is described in the video; no public Qwen 4 model card or price verifies it. | Treat the specs and promotion as unconfirmed until Alibaba posts terms and limits. |
| Open weights and local install | No official Qwen 4 download was located on Qwen's publisher page at this check. | Do not mistake third-party repositories named Qwen4 for an official release. Check license and file provenance on launch. |
| Qwen3.8-27B | Official weights and license; Ollama local package. | A real, installable baseline from an earlier generation; not evidence about Qwen 4 performance. |
What the Demos Are Good For
The strongest examples are the Roman Colosseum, a playable-looking village, an aquarium and SVG scenes. They show what a model-assisted build can look like. They do not reveal the full prompt, retries, editor tools, elapsed time or whether the same task succeeds repeatedly. The video's token-efficiency comparisons with GPT-6.1 Terra and DeepSeek V4.1 Flash are therefore creator-run observations, not a controlled benchmark.
The insulin-bioreactor simulation is particularly easy to overread. A coherent interactive interface and a plausible numerical output do not establish that the underlying biological model, yields or dose conversion are scientifically valid. Treat it as a software-demo prompt, not a medical or manufacturing result.
For your own evaluation, use the same task, time budget, tools and human review criteria across models. Save the generated files and inspect whether the app runs, whether controls work, and whether the model can fix a bug you deliberately introduce. A beautiful first screenshot is only the beginning of a useful coding test.
How to Run an Available Qwen Model Today
Qwen 4 cannot be installed from an official public checkpoint yet. The following steps install Qwen3.8-27B, a different, released model. The Ollama listing shows an approximately 18 GB download. Allow more than that on disk and in system/GPU memory for the runtime and context; the exact requirement varies with platform and requested context length. On a small laptop, choose a smaller official Qwen model rather than assuming this 27B package will be fast.
- Go to Ollama's official download page and install the Windows, macOS or Linux build appropriate for your machine. Prefer the installer page over copying an unreviewed shell command.
- Open a terminal or PowerShell after installation and check that
ollama --versionresponds. If it does not, finish the install and reopen the terminal. - Check the exact model tag, then run
ollama run qwen3.8:27b. Ollama downloads the package on first run and opens an interactive chat. - Ask it for a small, verifiable task, such as a single HTML page with one working interaction. Review the produced code and run it in a disposable folder before giving a coding agent filesystem access.
- Use
ollama listto confirm the local model is present. Ollama's quickstart and GPU guide help diagnose slow runs. A local inference process is not a guarantee that tools connected to an agent cannot send files elsewhere.
For the unquantized original rather than Ollama's packaged version, use only Qwen's model card and Hugging Face's download documentation. That route is substantially larger and needs a compatible inference stack. Do not label an unofficial conversion as Alibaba's original weights.
What to Do When Qwen 4 Actually Ships
First, find the checkpoint under Qwen's verified publisher account or an Alibaba announcement. Read the exact model card for license, modalities, context, memory, supported runtimes and download size. Then compare the released model on your own fixed tasks with the Qwen3.8 baseline. If only a hosted API appears, use the official API identifier and pricing page; hosted access is not the same as downloadable open weights.
You can try Arena Code now to compare anonymous coding outputs, but do not upload secrets or client code. Arena explains that prerelease models can appear under anonymous labels and that prompts support its evaluations. Its results can tell you which answer you prefer; they cannot by themselves certify a Qwen 4 model identity.
Three Companion Videos
These are the three additional episodes linked in the creator's description. They provide context for other model rumors; they are not independent confirmation of Qwen 4 specifications.
Separate product announcements from creator interpretation.
Useful reminder that an anonymous model's identity needs evidence.
Watch the claims, then look for the official release pages.
Social Links and Official Starting Points
| Source | Links | Use it for |
|---|---|---|
| Creator and demo accounts | World of AI · @loktar00 · @aijoey | Locating the creator's posts and code demos. A profile is not a substitute for a precise, verified post URL. |
| Other referenced accounts | @marcthecreatorr · @coffeecup2020 | Finding the demo threads described in the video; supplied status links were incomplete, so no specific post is attributed. |
| Anonymous testing | Arena Code · Arena methodology | Hands-on comparison and the limits of interpreting codenames. |
| Official release and install | Alibaba announcement · Qwen weights · Ollama baseline | Confirming what has actually shipped before installing or quoting specs. |
Video Chapters
Release and identity claims
| Time | Segment |
|---|---|
| 00:00 | Introduction |
| 01:41 | Release-date speculation |
| 02:07 | Preview outputs |
| 02:33 | Unverified specification leak |
| 03:14 | Alibaba teaser interpretation |
| 04:55 | Unknown preview size |
Coding and simulation demos
| Time | Segment |
|---|---|
| 05:11 | Frontend demo |
| 05:53 | Minecraft-style scene |
| 08:39 | Accessing Arena checkpoints |
| 09:22 | 3D Rome |
| 10:18 | Comparison with Terra |
| 11:09 | Bioreactor simulation |
| 12:07 | Reasoning settings |
| 13:05 | Comparison with DeepSeek |
| 13:32 | Shooter-game demo |
| 14:12 | Voxel test |
| 14:54 | Token-efficiency claim |
| 15:25 | Aquarium |
| 15:48 | SVG demo |