granite-4.0-h-350m-kor-16384

granite-4.0-h-350m-kor-16384

This model is a 18.95% smaller version of ibm-granite/granite-4.0-h-350m optimized for Korean language via vocabulary size reduction using the trimming method.
This trimmed model should perform similarly to the original model with only 16,384 tokens and a much smaller memory footprint. However, it may not perform well for other languages as tokens not commonly used in the selected languages were removed from the vocabulary.

Model Statistics

Mining Dataset Statistics

• Number of texts used for mining: 200,000 texts
• Dataset: lbourdois/fineweb-2-trimming

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda"
model_path = "alphaedge-ai/granite-4.0-h-350m-kor-16384"

tokenizer = AutoTokenizer.from_pretrained(model_path)
# drop device_map if running on CPU
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)
model.eval()

# change input text as desired
chat = [
    {"role": "user", "content": "Your prompt in Korean."},
]
chat = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

# tokenize the text
input_tokens = tokenizer(chat, return_tensors="pt").to(device)

# generate output tokens
output = model.generate(**input_tokens, max_new_tokens=100)

# decode output tokens into text
output = tokenizer.batch_decode(output)
print(output[0])

Citations

Granite 4.0

@misc{granite2025,
  author       = {IBM Research},
  title        = {Granite 4.0 Language Models},
  year         = {2025},
  howpublished = {https://github.com/ibm-granite/granite-4.0-language-models},
}

Trimming blog post

@misc{hf_blogpost_trimming,
      title={Introduction to Trimming}, 
      author={Loïck BOURDOIS and Tom AARSEN and Bram VANROY and Christopher AKIKI and Woojun JUNG and Manuel ROMERO and Prithiv SAKTHI},
      year={2026},
      url={https://huggingface.co/blog/lbourdois/introduction-to-trimming}, 
}