This is a fun side project, not a production-ready model. It can produce working KrunkScript for some basic requests, but quality is still poor. It may hallucinate APIs, mix in JavaScript syntax, or output broken code.

Simple LM Studio setup

Download and open LM studio.
Go to Model Search and download KrunkQwen-4B-Q4_K_M.gguf (search for KrunkQwen-4B-GGUF).
Load the model in Chat.
Set context length to 8192 or higher.
Open the system prompt from the GitHub repository: SYSTEM_PROMPT.md
Copy everything between BEGIN SYSTEM PROMPT and END SYSTEM PROMPT.
Paste it into LM Studio's System Prompt box.
Use:

text

Temperature: 0.1
Top P: 0.9
Max output tokens: 2000-4000

Test every generated script in the Krunker editor and report back if the script compiled or returned an error.

System prompt and project files

GitHub:

https://github.com/jajajiq/KrunkQwen-4B

Training summary

Base model: Qwen/Qwen3-4B-Instruct-2507
Method: Unsloth QLoRA
Epochs: 3
Training examples: 2,459
Quantization: Q4_K_M
DPO: not used

Limitations

KrunkQwen-4B may:

hallucinate nonexistent APIs;
output invalid KrunkScript;
mix in JavaScript syntax;
fail on complex or long requests;
produce incomplete or low-quality scripts.

The system prompt reduces mistakes but does not make the model reliable.

Future plans

A later version may use a much larger and cleaner public dataset, stronger and larger base models, better validation, more complete KrunkScript coverage, preference training, and stricter grammar checks.

Disclaimer

This is an independent community side project. It is not an official Krunker or KrunkScript project.

Jaj-Ajiq/KrunkQwen-4B-GGUF

KrunkQwen-4B