EXAONE Deep
LG AI's specialized inference model for solving complex problems step by step.
EXAONE Deep is an inference-specialized model developed by LG AI Research based on EXAONE 3.5. Beyond general text generation, it focuses on the ability of "deep thinking," which involves analyzing complex problems step-by-step and deriving logical conclusions. The inference process is explicitly separated using the `<thought>` tag, allowing for transparent verification of the logical path the model takes to reach a conclusion. In terms of performance, it is positioned as enhancing the inference capabilities of DeepSeek R1 70B with Korean language specialization, demonstrating superior results compared to similar models in mathematics, coding, and science benchmarks. In biotechnology research, the value of EXAONE Deep shines in complex experimental design and data interpretation. Similar to an experienced senior researcher who analyzes experimental results and reasons step-by-step, asking "The reason this value is unusual is...", the model formulates hypotheses, reviews refutations, and presents a final conclusion within the `<thought>` block. For example, if it detects an outlier in protein expression data, it lists possible causes (sample contamination, changes in experimental conditions, actual biological variation), evaluates each possibility, and then presents the most plausible explanation. It is available in a single 32B size and can be run with 19GB of VRAM. Since it maintains the same Korean language quality as EXAONE 3.5 while significantly enhancing inference depth, you can divide the roles by using EXAONE 3.5 for simple translation and summarization, and EXAONE Deep for complex analysis and judgment. It is officially supported in Ollama and, when deployed as an auxiliary engine for LocalRAG, generates highly reliable answers that include the inference process for in-depth questions.
💻 System Requirements
19GB (based on Q4 quantization)
19GB
⚡ Installation
### 4-1. Quick Start
```bash
ollama run exaone-deep:32b
```
### 4-2. Detailed Installation
```bash
# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh
# Download the model
ollama pull exaone-deep:32b
# HuggingFace Transformers
pip install transformers torch accelerate
```
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "LGAI-EXAONE/EXAONE-Deep-32B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
```🧬 Bio Use Cases
Case 1
Protein expression outlier analysis — When outliers are input from Western blot quantitative data, the `<thought>` block systematically evaluates possible causes such as sample contamination, loading errors, or biological variations. Finally, it presents the most likely cause and suggests additional experiments in Korean. (150 characters)
Case 2
Clinical trial statistical design review — The sample size calculation and statistical power analysis for a Phase II clinical trial are requested from EXAONE Deep. The model unfolds the process step-by-step, including effect size estimation, power calculation, and dropout rate correction, within the `<thought>` block, providing a review at the level of a statistician. (155 characters)
Case 3
Inference of complex drug interactions — Analyzes the expected pharmacokinetic interactions when three or more drugs are used in combination. It infers the CYP450 metabolic pathway, protein binding rate, and excretion competition step-by-step to classify the risk level. It uses accurate Korean clinical pharmacology terminology. (145 characters)
📝 Update Notes
No update notes yet.
🧪 Related Code of Life
No related Code of Life posts yet.