Qwen3-4B-Thinking-MiniMax-M2.1-Coder

Qwen3 4B Thinking x MiniMax M2.1 Code SFT

This model was trained on over 1,300 agentic "vibe coding" examples generated by MiniMax M2.1 with a large majority focused on extracting UI/UX design capabilities across different tech stacks.

For more info on how and what the model was trained on, please view the dataset card

How to run

Personally I use vllm on windows via wsl. Here is my command:

vllm serve TeichAI/Qwen3-4B-Thinking-MiniMax-M2.1-Code-Distill --reasoning-parser deepseek_r1 --enable-auto-tool-choice --tool-call-parser hermes --max-model-len 65536 --quantization bitsandbytes --override-generation-config '{"temperature": 0.6, "top_p": 0.95, "top_k": 20}'

Demo

Prompt: Make me a landing page for my bakery. we make cakes, cookies, brownies and everything else a normal bakery makes. I want it to look really nice

Needless to say I was impressed what this 4B model was capable of (especially at 4bit quant)

The site is overall very choppy and needs polishing, but please feel free to download the html file (in the repo) and give it a look with all it's wonky animations :)

---

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.