Can Open-Source AI Actually Produce Better UI Design than Claude?

I gave Kimi K3, GLM-5.3-Flash, and Claude Opus 5.5 the same real UI task inside our codebase. One produced something I would actually keep, and the gap was much bigger than I expected.

🔧 Get our free AI product and design skills:
https://designproject.io/skills-repo/?utm_source=youtube_open_source_ai_design&utm_medium=video&utm_campaign=open_source_ai_design&utm_content=description

💬 Join the TDP Community:
https://designproject.io/community/?utm_source=youtube_open_source_ai_design&utm_medium=video&utm_campaign=open_source_ai_design_community&utm_content=description

🚀 Explore a Product Sprint with TDP:
https://designproject.io/get-started/?utm_source=youtube_open_source_ai_design&utm_medium=video&utm_campaign=open_source_ai_design_sprint&utm_content=description

WATCH MORE LIKE THIS
🎥 We Gave 3 Designers the Same Prompt. Does Taste Still Matter? (Codex Test)
https://www.youtube.com/watch?v=zxRgj4PK4EU
🎥 I Used Claude Design on a Real Product Feature. My Honest Review
https://www.youtube.com/watch?v=m0sfGvVzGYs
🎥 I’ve Been a Claude Code Lover for a Year. Then I Tried Codex
https://www.youtube.com/watch?v=uswUyEnrSwg
🎥 7 AI Tools I Use to Ship Design Systems Faster
https://www.youtube.com/watch?v=2UhH5ZQtiGA

Open models are becoming cheaper, faster, and much easier to use. But for product teams, the interesting question is not how they perform on a benchmark. It is whether they can understand a real product, make a strong visual decision, and produce an interface you would actually ship.

I gave Kimi K3, GLM-5.3-Flash, and Claude Opus 5.5 the same prompt, the same codebase, and the same visual reference.

GLM understood the concept, but the copy and final composition needed work. Kimi created a similar transition, but the motion and hierarchy felt less polished. Claude produced the most refined interaction, although parts of the animation still needed another pass.

Open models are already useful for experimentation, coding, and lower-cost workflows. The right choice depends on the task, and every output still needs human judgment.

I also walk through how I access open models through Baseten, connect them to OpenCode inside Conductor, and switch between models without changing the rest of my workflow. This gives product teams more control over where inference runs and which provider handles their requests.

WHAT YOU’LL LEARN
- How Kimi K3, GLM-5.3-Flash, and Claude Opus 5.5 handle the same UI prompt
- Why I tested the models inside a real codebase instead of using a generic demo
- Where each of the three models performs well and where it falls short
- How model cost changes the decision for product teams
- What inference means when working with open models
- How to access open models through Baseten
- Open models are moving incredibly quickly, and I think they will become a much bigger part of product workflows. As these models evolve, product teams will have more choices around cost, infrastructure, customization, and how AI fits into their design and development process.

CONNECT WITH ME

🌐 Website: https://designproject.io/?utm_source=youtube_ai_product_workflow&utm_medium=video&utm_campaign=ai_product_workflow&utm_content=connect_website

💼 LinkedIn: https://www.linkedin.com/in/diannealter/

TIMESTAMPS
00:00 Can open source AI models produce great design?
02:32 Meet the models: GLM and Kimi
04:09 The design challenge: a scrolling landing page
07:41 Comparing costs: open source vs. Claude
09:15 Security and hosting with Baseten
10:53 How to connect Baseten to Conductor
14:47 Design results: GLM vs. Kimi vs. Claude

We at The Design Project help B2B SaaS product teams ship faster by combining AI tools with strategic design expertise. We've worked with 50+ teams, including companies acquired by Slack, Nvidia, and MrBeast, to cut product cycles in half using battle-tested AI workflows and design systems.