Qwen 4 max preview versus Deepseek 1 shot no harness Tiny World these are ugly but they're to illustrate 1 shot capabilities.
Qwen 4 result looks better easily however, this was after three tries with the first 2 failing.
Deepseeks is the first try.
Prompt below.
用 Qwen4 max preview 生成了做了一个秋季的白桦树林,全程只有一句提示词(我会放在下面),感觉还挺不错的啊,Qwen4 需要正确的引导就可以做出不错的审美作品呀🤔 提示词如下:
make a maple and birch forest at peak autumn, with trees and some nice animation, goal is to build a forest that has life to it. Without using any external assets,all in code.
@Alibaba_Qwen@QwenAIFans
A sprinkle of DeepSeek in a sea of Qwen4-max-preview posts.
Here is a really good looking isometric village diorama: medina; created by deepseek-v4.1-flash-1017 without external assets.
All in code, single shot result, not with any harness and no tool calls.
Total tokens 180,105 · 13m 40s · 219.5 tok/s
I used Qwen4-Max-Preview in Qoder to build a hyper-realistic, NFS-inspired racing game. 🏎🟀
The goal was to create an immersive driving experience with realistic visuals, a detailed racing environment, and the feel of an actual racing game.
Pretty fun to see the idea turn into a playable experience and explore how much AI can handle in game development.
Qwen4-Max-Preview is now available in Qoder, Alibaba's coding agent, with a limited-time 80% discount.
If you've been curious about what it can build, give it a try!
#Qoder #Qwen4MaxPreview #GameDevelopment
Qwen4-Max made a deep ocean full of glowing jellyfish in Three.js. Same prompt in 2 harnesses.
Top: DeepSeek Harness. Bottom: Qwen Code. Both on xhigh.
The Qwen Code version flashed black boxes for 1 frame at a time, the DSH one blew the screen white on click. 1 bug report each and both fixed it, so the model can debug its own shaders when you point at the problem.
Tokens, DSH: 11.22M input, 10.80M cached, 110K output
Tokens, Qwen Code: 10.84M input, 10.18M cached, 121K output
Prompt:
Create a deep-ocean scene in Three.js, in a single HTML file: a swarm of bioluminescent jellyfish drifting through dark blue water. Each jellyfish has a soft, pulsing bell made with a vertex shader and long tentacles simulated with verlet physics that trail and sway as it swims. Add glowing light patterns that ripple across the bells, volumetric light shafts from the surface far above, drifting marine snow particles, depth fog and bloom. Moving the mouse creates a gentle current that pushes the jellyfish and bends their tentacles; clicking makes the nearest ones flash and dart away. Keep it smooth at 60 fps. When you are done, open the page in the browser, look at screenshots and fix what looks wrong.
Qwen 4 Max Preview
This is one of the best outputs I've gotten from it after running several tests
it's good but not that good it's not better than GPT 6 Sol or opus 5
I was expecting a much bigger upgrade especially after their claims about RSI
I've been curious about what Qwen4 could do with 3D, so when Qoder rolled out Qwen4-Max-Preview, I decided to put it to the test. It's currently available at a limited-time 80% off.
I make 3D teardown videos for a living, so I gave it a real product to work with.
One prompt in Qoder, and 30 minutes later, I had this interactive 3D Mac mini. It flips over to reveal the power button, opens up, and zooms right into the M4 chip.
I was especially impressed by how accurately it recreated the ports and internal details. Definitely curious to see how far I can push it with more complex builds.
I've been watching everyone guess what Qoduck was.
Turns out it's Qwen4-Max-Preview, inside Qoder.
So I gave it a task that's hard to fake: build a realistic old pirate ship on the ocean, in 3D.
It built "The Salt Widow."
→ Three-masted galleon with a raised stern and rope rigging
→ Tattered sails that billow in the wind
→ Glowing lanterns along the hull
→ A sunset sky with an animated sea
→ A Jolly Roger you can toggle on and off
→ A "Master's Instruments" panel: sea state, wind speed, chop, sail trim, and sail wear, all live
→ Sound, a compass, and a "Frame the Ship" camera button
All procedural. No external assets.
I did not expect the first result to look like this.
(80% off in Qoder for a limited time, if you want to try it.)
Building a mini game with an AI coding agent? 🕹️
I've been curious about what Qwen4 can actually do, so I decided to try Qwen4-Max-Preview in Alibaba's coding agent Qoder.
I used it to build a mini game from scratch, turning a simple idea into something interactive that I could actually play.
The final result came together nicely, with a clean interface, engaging gameplay, and a smooth overall experience.
It's pretty cool to see how far you can get with just an idea and an AI coding agent, without having to handle every technical detail yourself.
Qwen4-Max-Preview is now available in Qoder with a limited-time 80% discount.
#Qoder #Qwen4MaxPreview #AI #Coding
Hugging Face is now the best place to host and share 3D assets 🧊🤗
Drop in .glb, .ply, .stl, or splat files and get a 3D gallery automatically. There are thousands of 3D models already in public datasets!
350+ datasets to explore: https://t.co/yQv82qMWL9
Docs: https://t.co/cyTYMSpJos
做 3D 游戏的开发者看过来!
这里有超过 50,000 份免费、开源的游戏素材,包括 3D 角色、怪物、机甲、建筑、场景、贴图、材质和音效,做游戏需要的各种资源都可以在这里找。
Patch 刚刚推出了 Patch Toolbox,把这些游戏素材整理到一个资源库里。可以直接按关键词搜索,预览模型,查看授权信息,再把需要的资源用到自己的游戏项目中。
更有意思的是,它还支持 CLI 和 MCP,可以接入 Claude Code 等 AI 编程工具。
也就是说,以后我们让 AI 开发游戏时,不仅可以让它编写代码、设计玩法,还可以让它直接搜索合适的 3D 模型、角色和场景素材,整合进游戏项目。
我觉得这个方向很有潜力。
现在用 AI 写游戏代码已经越来越快,但美术资源仍然是独立开发者的一大难题。尤其是一个人开发游戏,想把角色、建筑、场景、动画和音效全部做好,工作量相当大。
其实很多游戏根本没必要从零制作每一个模型。
如果 AI 能直接调用成熟的素材库,再配合自动编程、关卡设计和场景搭建,一个人开发一款画面漂亮、玩法完整的 3D 小游戏,成本和时间都有机会大幅下降。
以后独立游戏开发的竞争,可能越来越取决于创意和玩法,而不是谁能投入更多时间制作基础素材。
项目地址:https://t.co/TZ03Ii129p