EXAONE 3.5
LG's state-of-the-art Korean NLP, a dual-language open-source LLM.
EXAONE 3.5 is a bilingual (Korean-English) large language model developed by LG AI Research, specializing in Korean. It boasts top-tier performance among existing open-source models in Korean NLP. Available in two sizes, 7.8B and 32B, the 32B model, in particular, demonstrates GPT-4-level quality in Korean benchmarks. It generates remarkably natural results compared to foreign models in tasks such as Korean paper summarization, translation, and document classification, making it virtually an essential tool for Korean biotechnology researchers. EXAONE incorporates decades of Korean data and domain knowledge accumulated by LG. It accurately uses particles and suffixes, and the inclusion of Korean equivalents for technical terms feels natural, as if written by a researcher whose native language is Korean. When translating English-language academic papers into Korean, it excels at maintaining the academic context, translating terms like "Protein Folding" appropriately. Based on the 32B model, it achieves approximately 10.6 tok/s on a GPU and can be run with 19GB of VRAM, making it executable even on consumer GPUs. The 7.8B model is a lightweight version, suitable for chatbots requiring real-time responses or for fast document classification. The 32B model is used for in-depth analysis of long documents and tasks requiring complex reasoning. It is officially supported in Ollama and can also be directly loaded using Hugging Face Transformers. When used as an automatic Korean description generation (generate-desc) engine in the AI Tools pipeline, it can produce high-quality Korean catalogs.
๐ป System Requirements
7.8B โ 6GB / 32B โ 19GB (based on Q4 quantization)
7.8B โ 5GB / 32B โ 19GB
โก Installation
### 4-1. Quick Start
```bash
# 7.8B model (lightweight)
ollama run exaone3.5:7.8b
# 32B model (high-quality Korean)
ollama run exaone3.5:32b
```
### 4-2. Detailed Installation
```bash
# Using HuggingFace Transformers
pip install transformers torch accelerate
```
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "LGAI-EXAONE/EXAONE-3.5-32B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
```๐งฌ Bio Use Cases
Case 1
Automated Korean Summarization of Scientific Papers โ Summarize the abstracts of 100 PubMed English-language papers into Korean using EXAONE 3.5 32B. Automatically apply academic terms in the format of "Protein Folding." Assessed as equal to or better than GPT-4's Korean summarization in terms of summary quality. (150 characters)
Case 2
Korean Generation of Clinical Trial Reports โ Automatically generate a draft of a Korean clinical trial report by inputting English CRF (Case Report Form) data. Korean transliteration of medical terms (e.g., Adverse Event โ ์ด์๋ฐ์) is consistent with KFDA guidelines. Report writing time reduced by 70%. (155 characters)
Case 3
Biotechnology Patent Classification โ Classify 500 Korean patent documents in real-time using a 7.8B model. Accuracy of 94.2% for the three major categories based on IPC codes: A61K (Pharmaceuticals) / C12N (Genetic Engineering) / G01N (Analysis). Lightweight model processes each document in under 2 seconds. (135 characters)
๐ Update Notes
No update notes yet.
๐งช Related Code of Life
No related Code of Life posts yet.