Qwen4-Max Preview Tested: Claw Machines, Ink Paintings & 3D Worlds Show Strong Token Efficiency
The author benchmarks Qwen4-Max Preview on five creative coding tasks—including a physics-based claw machine, a procedural Chinese ink painting engine, a Three.js League of Legends map, and a time-collapse visualization—finding exceptional token efficiency and engineering quality despite slow generation speeds of ~33 tokens/s.
Overview
Alibaba's Qwen4-Max Preview (Qwen4-Max-Preview) launched on Token Plan at a discounted rate (19.5 CNY for 20,000 credits, 0.1x multiplier). The author tested it via Qoder (CN Pro 19.9 CNY) across five tasks, highlighting three upgraded capabilities: deeper knowledge understanding, more three-dimensional world perception, and stronger coding and agentic skills on real tasks.
1. Claw Machine Challenge: Physics, State Machines & Frontend Animation
Reproducing a prior benchmark (DeepSeek, Kimi K3.1, Opus 5.5, GPT-6 token comparison), the task tests physical interaction, state machines, and frontend effects. Qwen4-Max took ~75 minutes (slow output ~33 tokens/s) but delivered a solid implementation: mechanical force feedback, rope pendulum swing, doll collision, and scoring logic. The author rates it on par with Kimi K3.1.
Token efficiency comparison (estimated tokens via Qoder credits):
Codex : 8 min, 3.5M total (97% cached), 28K output
Claude Code : 17 min, 12.15M total (96% cached), 130K output
Gemini : 21 min, 13.94M total (96% cached), 81K output
Kimi Code : 36 min, 9.5M total (98% cached), 59K output
DSH : 46 min, 46.44M total (99% cached), 281K output
MiniMax Code : 68 min, 40M total (98% cached), 198K output
Qwen4-Max : 75 min, 2.69M total (95% cached), 55K output
Qwen4-Max shows the lowest total token consumption among tested models, with large action granularity and minimal error-driven rewrites, yielding high effective code ratio.
2. Procedural Ink Painting "双燕": Hand‑Written Physics Engine & Self‑Check
Prompt: Python (numpy + PIL) painting in Wu Guanzhong style, no image models, no render‑then‑trace. Qwen4-Max built a custom physics ink simulation engine ( ink_engine.py) modeling:
Multi‑scale absorbency of rice‑paper fibers
Water diffusion and edge "water‑line" deposition
"Flying white" (干笔飞白) from dry brush
Brush as dozens of individual bristles, distinguishing goat‑hair softness vs wolf‑hair stiffness
Five ink tones: 焦 (scorched), 浓 (dense), 重 (heavy), 淡 (light), 清 (clear)
Output: 1,000+ lines, 4,019 precise brush strokes. A second piece "双燕" exhibited strong artistic mood.
Self‑check logic : The model wrote numerical verification—pixel brightness stats (blank ratio ~64%, white‑wall brightness 84–100%, dense ink ≤5.8%) and ASCII contour scanning. It detected swallows overlapping the dark window frame and autonomously translated coordinates onto the clean white wall.
Compared to Opus 5.5 on the same prompt, Qwen4-Max still trails in relaxed mood and artistic brushwork.
3. Three.js Summoner's Rift: 1,263‑Line Holographic Sandbox
Three.js implementation of League of Legends 3D map with:
Three lanes, towers, base, dragon/baron pits, curved river via parametric sine equations
Minion waves auto‑spawn, fight, push towers; tower destruction with debris collapse animation
Jungle monsters with respawn countdown
14 persistent UI controls: day/night slider (fireflies at night), tower crystal breathing glow, clickable minimap with camera jump
Versus GPT‑6 Astra, Qwen4‑Max‑Preview shows richer dynamic interaction, combat choreography, and sandbox vitality.
4. "End of Time": Cosmic Aesthetic & Gravitational Collapse
Using a prompt from Kazek (testing GPT‑6.2 Sol), Qwen4‑Max created a Three.js interactive page "End of Time": black‑gold abyss with layered golden star‑track rings. Highlight: "Time Zero" interaction—clicking the central dial collapses swarming time particles into a singularity, then erupts in a blinding white flash resetting the scene, delivering strong visual impact.
5. Overseas Community Demos
@PokeBenchTV: Sunny ship & ocean scene replication; model added unrequested sliders, sound, compass, quality settings; ~5.4M tokens
@eddiboi: "Lightning in a bottle"
@hollaluccaxo: Dune racing
Conclusion
Qwen4‑Max‑Preview excels in token thrift and engineering rigor across diverse creative coding tasks. Main drawback: slow generation (~33 tokens/s), likely due to preview‑phase concurrency scheduling and compute quota limits. The author believes Alibaba has all ingredients for a comeback—compute, cloud infrastructure, national‑scale apps, talent, global open‑source pull—and hopes for an official Qwen4 release matching the caliber of Google's Argon/Carbon models.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
