Open Source AI Model Comparison | GPT-OSS | DeepSeek | Qwen | Kimi-K2

Описание: Comparison of OpenAI's new open source AI model GPT‑OSS 120B compared to other frontier open source/open weight AI models with visual simulation coding challenges.

Models
GPT‑OSS 120B
DeepSeek R1
Kimi-K2
Llama Maverick
Qwen3 Coder
MiniMax-M1
GLM-4.5

0:00 Spinning Hexagons
1:02 Double Pendulum
1:57 Mini Planet

Testing Structure

The examples shown are the best of 4 tries from each model for each prompt. Each model had identical prompts except for Kimi-K2 and DeepSeek which required some prompt additions to get valid output.

All models used maximum reasoning effort.

Insights From Testing

Models with most consistent quality on these tasks:
1. Qwen3 Coder
2. GPT‑OSS 120B

DeepSeek was very high quality, but insisted on adding user input controls even though the prompt code requirements stated this was not allowed.

Models with low consistency on the prompts:
Kimi-K2
Llama Maverick
MiniMax-M1

Kimi-K2 could not produce valid code for the mini-planet task until I told it to write out a detailed plan in an html comment near the top of it's code, then it produced the output shown.

Qwen: Qwen3 235B A22B Thinking was difficult to use on the providers I tried, API calls kept timing out with huge thinking blocks, and samples I did get weren't as good as Qwen3 Coder. Nemotron Ultra was also tested, but the results weren't functional enough to include.

The core prompts are shown in the video, but they also included some additional coding requirements at the end that allowed me to automatically record and export the videos in bulk. Requirements are only partially shown here as I can't include code examples in this description.

HTML OUTPUT REQUIREMENTS

Output a complete HTML file containing your visualization.

*Required:*
Use `canvas.width` and `canvas.height` for dimensions (framework sets these)
No CSS width/height on canvas.=
No user input or text elements allowed