Alibaba’s Tongyi Qianwen had only just been rumored to be about to launch Qwen4 when word recently emerged that internal testing has begun: the leaker relayed the views of fellow members in the internal test group, saying its front-end coding ability sits roughly between Opus 5.5 and Astra, its aesthetic performance beats DeepSeek’s gray-test version, and the highly anticipated Qwen4-27B may also arrive ahead of schedule. The Apsara Conference has confirmed the Qwen4 family naming and the 27B open-source plan, but officials have not yet announced the timeline or specs. Below is a rundown of the backstory of this leak, along with a breakdown of what those demos are actually showing off.
Internal beta screenshots leaked; source: ModelScope Community.
On the morning of September 28 Taipei time, the X account AJ (@ItsmeAjayKV) posted early internal test results for Qwen4, saying that after testing, members of the internal test group believed its front-end programming was roughly at the level of Opus 5.5 and Astra, “possibly slightly lower,” but that its aesthetics were better and more soulful than DeepSeek 0820’s gray test, and that Qwen4-27B would also be out soon and they couldn’t wait for October. This timing was only 6 days after the Apsara Conference officially teased Qwen4, meaning that no sooner had the official side made its announcement than the internal test results leaked right after.
Rejoice Qwen lovers!
Early qwen4 internal testing results are starting to appear!
They are testing its frontend coding and found it roughly in the Opus 5.5/Astra range, with better visual taste than the DeepSeek 0820
Soon we will start seeing Qwen4-27B out there as well,… pic.twitter.com/VRUNSxVV1U
— AJ (@ItsmeAjayKV) September 27, 2026
The original source of the image attached to this tweet is a beta test post titled “qwen4 Beta Test” by ModelScope community user “空什么色”, which has accumulated 367 likes and 64 comments.]The original text explicitly states:
I’m not exaggerating when I say the closed beta results are already dynasty-tier. The group folks have tested the frontend, and I can only say it’s insanely strong. The frontend is at the level of opus5.5 and astra—maybe slightly lower, but it already completely crushes every domestic model by a wide margin. It’s on the same level as deepseek0820 in gray testing, but with better taste and more soul.
The post includes several images: one shows a Chinese palace complex scene, and another shows a pelican doing a backflip on a bicycle. The beta tester emphasized that the pelican video “has no bicycle proportion issues or motion issues, and the backflip motion is perfect.” The original text specifically notes that the pelican is a moving video rather than a static image, meaning this round of beta testing demonstrations spans both front-end programming and image generation.
AJ then added two detailed screenshots that went viral, with watermarks from Baidu Tieba’s “deepseek” bar: one is a “Voxel Sangharama” voxel Chinese Buddhist temple scene implemented in Three.js, with on-screen stats showing 128,293 voxels drawn, only 5 draw calls, and a build time of 142ms, plus four lighting toggles for dawn, noon, dusk, and night; the other is a Minecraft-style Chinese palace complex.
What the demo is showing off: voxel Buddhist temple and backflipping pelican
First, to explain the difficulty of the voxel scene: “Voxel Sangharama” is a Chinese-style Buddhist temple built from small cubes, and the key is that it is rendered in real time with Three.js. The stats in the upper-left corner of the screen are the report card: over 120,000 voxels compressed into 5 draw calls, meaning the model knows how to batch-merge geometry instead of naively producing hundreds of thousands of draw commands; a construction time of 142ms is also within the usable range. The four time-of-day lighting switches are achieved by parameterizing the scene and writing it as a switchable program structure, rather than a hardcoded static image. This kind of challenge tests the ability to translate design aesthetics into executable code, which is exactly the concrete manifestation of “better aesthetics” in the leak.
The pelican doing a backflip on a bicycle is another thread. This kind of absurd-action video is often used to test the physical consistency of image generation: whether the bicycle’s proportions collapse, whether the feathers and frame stick together, and whether the center of gravity is off during the backflip. Internal testers’ comments on this segment centered on “no proportion issues or motion issues at all.” Incidentally, the “deepseek0820 gray test” mentioned in the original text did not specify version details; from context, it is inferred to be the DeepSeek V4.1 Flash gray-test version from around late August.
Official progress: Qwen4 is in training, with 27B to be open-sourced first.
At the Yunqi Conference on September 22, Alibaba confirmed that Qwen4 is in training and uses an all-new architecture. The family is named Qwen4-Max, Qwen4-Plus, Qwen4-Flash, and Qwen4-27B. The slide only said “Coming Soon” and gave no date. On the open-source side, Liu Dayiheng, head of Qwen LLM, confirmed that Qwen4-27B will be prioritized. On HuggingFace, Qwen3.8-27B has already become the platform’s most popular open-source model, so the successor Qwen4-27B is naturally drawing considerable attention.
On the timing front, Chinese media, citing leaks, point to an October release, consistent with AJ’s tweet “can’t wait for October,” while officials have remained tight-lipped about the date. As for the rumored 5 to 10T parameter scale circulating externally, that refers to expansion plans for future generations such as Qwen4.5 and Qwen5 and has nothing to do with Qwen4 itself.
How should we view this wave of leaks?
Right now, the entire story revolves around internal-test screenshots and community leaks; there is no official Qwen endorsement of specs, benchmark scores, or timeline. The claim of being at “Opus 5.5 and Astra level” comes from subjective comments by internal testers, and the person who made the post also added “may be slightly lower.” The two detailed screenshots being widely circulated bear Tieba watermarks and have passed through several rounds of reposting. What stands up better at the moment is three things: the Qwen4 family naming, the 27B open-source schedule, and the architecture preview already being open. As for the “dynasty-level” internal-test performance, once the October version arrives and the same set of questions is used to retest, the answer will naturally become clear.
Source: KOCPC Chinese