Qwen4 released
Read the Portuguese translation →
Author: Data Science in your pocket · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Qwen4 released · Channel: Data Science in your pocket · 04:15 · Watch on YouTube
Topics: AI evaluation and benchmarks, AI content generation, AI models, Local and open-source models
Summary
Section titled “Summary”The video discusses an early 3D helicopter demo attributed to Qwen 4, which Alibaba says is still in training and has not yet been formally released. It compares the demo with Kimi K 3.1 and reviews Qwen’s recent model sizes, public benchmark figures, and a preview of the architecture. The presenter cautions that a polished demo does not establish reliability across real-world tasks, while noting uncertainty around Qwen 4’s tiers, pricing, and licensing. Viewers are advised to look for model documentation, stable APIs, repeated independent benchmarks, and usable open-weight terms when the model launches.
Key points
Section titled “Key points”- Alibaba said Qwen 4 was in training, with no model card, API, or pricing announced at the time described.
- A helicopter-building demo attributed to Qwen 4 reportedly took roughly 15 to 20 minutes, matching Kimi K 3.1’s timeframe.
- Qwen 4’s size has not been published; the video cites Qwen 3.8 Max as the closest public proxy.
- The cited Qwen 3.8 Max figures include 92.6 on GPQA Diamond, 93 on Paper Bench, 86.1 on OS World, and a 1 million-token context.
- Qwen 3.8 Flash Next is described as an official preview of the Qwen 4 architecture, using a large memory table and sparse active parameters per token.
- The demo shows coordinated geometry, rendering, controls, and animation, but does not demonstrate reliability across many prompts.
- Qwen 4 is expected to have Max, Plus, Flash, and 27B tiers; the video says its licensing terms remain unknown.
- The presenter recommends checking the model card and pricing, API identifiers, repeatable independent benchmarks, and open-weight terms.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | An early Qwen 4 demo | The video introduces a detailed helicopter-building demo and frames its appearance before Qwen 4’s official release as the central story. |
| 01:35 | The architecture and helicopter demo | The presenter outlines the Qwen 3.8 Flash Next architecture preview and describes the side-by-side 3D demo with Kimi K 3.1. |
| 02:39 | A demo is not a reliability test | The video argues that a successful project does not prove reliability across many prompts, then discusses expected tiers and licensing uncertainty. |
| 03:44 | What to watch when Qwen 4 ships | The presenter names four release details to check and concludes that the next contest is reliability rather than demos. |
Quotes
Section titled “Quotes”The demo is here before the product. — 00:00
Those are the numbers Qwen 4 has to beat, not the helicopter. — 01:03
But an impressive demo is not a benchmark. — 02:39
The demo race is over. The reliability race is next, and that one has no leaderboard yet. — 04:15
Concepts and entities
Section titled “Concepts and entities”Qwen4, Alibaba, Kimi K3.1, Qwen3.8 Max, Qwen3.8 Flash Next, GPQA Diamond, PaperBench, OSWorld, Code Arena, Apache License 2.0
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 One prompt, a working 3D helicopter, takeoff, landing, assembly, exploded views, x-ray, and wireframe. Built in roughly 15 to 20 minutes. By a model Alibaba still has not launched. This is Qwen 4 showing up early. On September 22nd, 2026, Alibaba said Qwen 4 was in training. No model card, no API, no price. One month later, it is matching Kim i K 3.1’s helicopter render
00:33 prompt for prompt. That gap is the story. The demo is here before the product. Qwen 4 sits on a very fast ladder. Qwen 3.5 started at 397 billion parameters. Qwen 3.8 max jumped to 2.4 trillion with 95 billion active. Qwen 4’s size has not been published. The roadmap only says Qwen 4.5 and Qwen 5 will scale to 5 to 10 trillion. Everything in between is
01:03 inference. The closest public proxy is Qwen 3.8 max. It scores 92.6 on GPQA Diamond, 93 on Paper Bench, and 86.1 on OS World. It ranks fourth on Code Arena. 1 million token context. $2 input, $6 output. Those are the numbers Qwen 4 has to beat, not the helicopter.
01:35 Alibaba has already shown the blueprint. Qwen 3.8 Flash Next is an official preview of the Qwen 4 architecture. 125 billion parameters plus a 51 billion parameter Ingram memory table with only 6 billion active per token. 512 experts, 10 routed plus one shared. Hybrid linear attention and a native 262K context that extends to 1 million. The trick is capacity without proportional compute.
02:08 Then there is the demo. The same prompt, takeoff and landing, full assembly, exploded view, X-ray, wireframe. Lumina reports both Qwen 4 and Kimi K 3.1 finishing in roughly 15 to 20 minutes. Not a benchmark, a complete 3D application built end to end. Side by side, the outputs are close. Both land in the same 15 to 20-minute window. Both deliver the same
02:39 interaction model, takeoff, assembly, exploded views, and inspection modes. The gap between the two labs is now measured in minutes, not generations. But an impressive demo is not a benchmark. It proves the model can coordinate geometry, rendering, controls, and animation in one project. It does not prove reliability across 10,000 prompts. The right question is not, can it build a helicopter? It is,
03:09 can it build your app then fix it when it breaks? Qwen 4 is expected to arrive as four tiers, Max, Plus, Flash, and a 27B. The 27B is the one open-weight developers are watching. Qwen’s recent history is mixed. The last 27B shipped under Apache 2.0. Qwen 3.8 Max came with a custom license and a $50 million revenue trigger. Qwen 4’s terms are still unknown.
03:44 When Qwen 4 ships, ignore the launch video. Watch four things, a model card with real context and pricing, stable API identifiers, independent benchmarks run more than once, and open weights with terms you can live with. The helicopter got our attention. Those four things decide whether it belongs in production. Quinn 4 is not a released model yet. It is a warning shot, a frontier class demo from a model still in training, arriving
04:15 in the same 15 to 20-minute window as Kimi K 3.1. The demo race is over. The reliability race is next, and that one has no leaderboard yet.