Arcada Labs Code Categories Arena Leaderboard
The latest AI design-code model leaderboard based on Arcada Labs Code Categories Arena anonymous user voting. Focused on Website, UI components, game development, and data visualization code generation.
Top Model
GLM 5.2
Top Score
1352.00
Model Count
136
Data version
2026年07月12日
Data source: Arcada Labs
Ranking Table
| Rank | Model | Score | 95% CI | Votes | Organization | License |
|---|---|---|---|---|---|---|
| GLM 5.2Zhipu AI | 1352.00 | +/-7.9 | 8,533 | Zhipu AI | Open Source | |
| 9 |  Kimi K2.6Moonshot AI | 1320.00 | +/-4.9 | 23,919 | Moonshot AI | Open Source |
| 14 | Kimi K2.7 CodeMoonshot AI | 1300.00 | +/-8.2 | 7,614 | Moonshot AI | Open Source |
| 16 |  MiniMax M3MiniMax | 1297.00 | +/-6.5 | 12,553 | MiniMax | Open Source |
| 23 |  DeepSeek-V4-ProDeepSeek-AI | 1281.00 | +/-5.5 | 17,824 | DeepSeek-AI | Open Source |
| 24 | Kimi K2.5 (thinking)Moonshot AI | 1279.00 | +/-4 | 39,455 | Moonshot AI | Open Source |
| 26 |  MiniMax-M2.7MiniMaxAI | 1275.00 | +/-4.4 | 31,074 | MiniMaxAI | Open Source |
| 35 | MiniMax M2.5MiniMaxAI | 1249.00 | +/-6.7 | 11,504 | MiniMaxAI | Open Source |
| 36 | DeepSeek-V4-FlashDeepSeek-AI | 1248.00 | +/-4.7 | 25,437 | DeepSeek-AI | Open Source |
| 39 | M2.1MiniMaxAI | 1232.00 | +/-5.1 | 20,805 | MiniMaxAI | Open Source |
| 48 | DeepSeek-V3.1 (thinking)DeepSeek-AI | 1217.00 | +/-5.7 | 16,258 | DeepSeek-AI | Open Source |
| 50 | DeepSeek V3.2-ExpDeepSeek-AI | 1213.00 | +/-5.3 | 19,490 | DeepSeek-AI | Open Source |
| 55 |  Step 3.7 FlashStepFun | 1210.00 | +/-6.3 | 13,365 | StepFun | Open Source |
| 62 | DeepSeek V3.2DeepSeek-AI | 1206.00 | +/-4.4 | 29,394 | DeepSeek-AI | Open Source |
| 65 |  Hy3Tencent | 1199.00 | +/-20.5 | 1,202 | Tencent | Open Source |
| 76 | DeepSeek-R1-0528DeepSeek-AI | 1181.00 | +/-5.4 | 17,944 | DeepSeek-AI | Open Source |
| 79 | MiniMax M2MiniMaxAI | 1177.00 | +/-6.9 | 10,828 | MiniMaxAI | Open Source |
| 87 | DeepSeek-V3.1DeepSeek-AI | 1154.00 | +/-5.1 | 20,278 | DeepSeek-AI | Open Source |
| 89 | DeepSeek-V3-0324DeepSeek-AI | 1151.00 | +/-5.3 | 19,257 | DeepSeek-AI | Open Source |
| 93 | Kimi K2 0905Moonshot AI | 1140.00 | +/-17.9 | 1,504 | Moonshot AI | Open Source |
| 98 | Kimi K2 Turbo PreviewMoonshot AI | 1126.00 | +/-15.2 | 2,094 | Moonshot AI | Open Source |
| 109 | Kimi K2Moonshot AI | 1076.00 | +/-19.4 | 1,352 | Moonshot AI | Open Source |
| 111 |  Qwen3-235B-A22B-Thinking-2507Alibaba | 1075.00 | +/-9.1 | 6,169 | Alibaba | Open Source |
Data is for reference only. Official sources are authoritative. Click model names to view DataLearner model profiles.
About This Leaderboard
This leaderboard uses data from Design Arena developed by Arcada Labs, a Y Combinator-backed platform for anonymous head-to-head evaluation of AI design-code generation.
Unlike LMArena's general text and coding evaluations, Design Arena's code leaderboard focuses on the ability to generate front-end code with visual output. Tasks include Website, UI components, game development, data visualization, SVG, web apps, mobile, and related subcategories.
This page shows the Code Categories aggregate ranking. Votes across subcategories are pooled and scored with a Bradley-Terry model. Votes are counted equally rather than category-weighted, so categories with more votes can influence the aggregate more.
FAQ
What is Arcada Labs Code Categories Arena?
Arcada Labs Code Categories Arena is an anonymous evaluation platform focused on AI design-code generation. It covers categories such as websites, UI components, game development, and data visualization, then aggregates votes into an overall ranking.
How is Arcada Code Arena different from LMArena Coding Arena?
LMArena Coding Arena focuses on general programming tasks such as code generation, debugging, and algorithms. Arcada Code Arena focuses on visual front-end outputs such as HTML pages, interactive UI components, charts, SVG, and prototypes.
What is the ranking methodology?
Arcada Labs pools raw votes from code subcategories and fits a Bradley-Terry model. Votes are equal rather than category-weighted, so higher-volume categories can influence the aggregate more.
Which model types perform best for design-code tasks?
Large models with strong visual reasoning and front-end coding ability tend to do well. Specialized UI and code-generation models can also perform strongly when tasks emphasize layout, interaction, and visual polish.