CarrotAI/Rabbit3-Ko-4B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:May 16, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

CarrotAI/Rabbit3-Ko-4B is a 4 billion parameter instruction-tuned large language model developed by CarrotAI, based on Qwen/Qwen3-4B. Fine-tuned using both Korean and English datasets, it is optimized for bilingual performance, particularly excelling in Korean language tasks. This model is designed for general instruction-following applications, demonstrating strong capabilities in mathematical reasoning and common sense understanding.

Loading preview...

CarrotAI/Rabbit3-Ko-4B Overview

CarrotAI/Rabbit3-Ko-4B is a 4 billion parameter instruction-tuned large language model developed by CarrotAI. Built upon the Qwen/Qwen3-4B base model, it has been extensively fine-tuned using a combination of Korean and English datasets, making it a strong performer in bilingual contexts, with a particular focus on Korean language processing.

Key Capabilities

  • Bilingual Proficiency: Optimized for both Korean and English, enabling effective communication and instruction following in both languages.
  • Instruction Following: Designed to accurately interpret and execute user instructions across various tasks.
  • Mathematical Reasoning: Achieves a gsm8k score of 0.8400 (5-shot exact match) and hrm8k_math score of 0.5539, indicating solid performance in mathematical problem-solving.
  • Common Sense & General Knowledge: Demonstrates capabilities in general knowledge benchmarks like kmmlu_direct (0.5212) and kobest (0.7768 accuracy).

Good For

  • Applications requiring robust Korean language understanding and generation.
  • Instruction-following tasks in bilingual (Korean/English) environments.
  • Use cases involving mathematical reasoning and general knowledge queries.

This model is intended for general use and can be deployed using frameworks like sglang or vllm for OpenAI-compatible API endpoints.

Finetunes
1 models