EXAONE-Deep-32B
| Property | Value |
|---|---|
| Parameter Count | 30.95B (without embeddings) |
| Context Length | 32,768 tokens |
| Architecture | 64 layers, GQA with 40 Q-heads and 8 KV-heads |
| License | EXAONE AI Model License Agreement 1.1 - NC |
| Model Hub | Hugging Face |
What is EXAONE-Deep-32B?
EXAONE-Deep-32B is a large language model developed by LG AI Research, specifically designed for advanced reasoning tasks. It represents the largest variant in the EXAONE Deep series, showcasing exceptional performance across mathematical reasoning, coding, and complex problem-solving scenarios. The model utilizes a sophisticated architecture with Grouped-Query Attention (GQA) and supports an impressive 32K token context window.
Implementation Details
The model architecture features 64 layers with a unique attention mechanism using 40 query heads and 8 key-value heads in its GQA implementation. It employs a vocabulary size of 102,400 tokens and is optimized for bfloat16 precision. The model can be deployed through various frameworks including TensorRT-LLM, vLLM, SGLang, and llama.cpp.
- Advanced reasoning capabilities demonstrated through high performance on MATH-500, AIME, and CSAT benchmarks
- Optimized for mathematical and coding tasks with specialized prompting structure
- Supports streaming inference with temperature and top-p sampling controls
- Available in multiple quantized versions using AWQ and GGUF formats
Core Capabilities
- 95.7% accuracy on MATH-500 benchmark
- 72.1% pass rate on AIME 2024 problems
- 94.5% accuracy on CSAT Math 2025
- Strong performance in coding tasks with 59.5% accuracy on Live Code Bench
- Competitive reasoning capabilities against larger models including DeepSeek-R1
Frequently Asked Questions
Q: What makes this model unique?
EXAONE-Deep-32B stands out for its specialized reasoning capabilities and optimized architecture for mathematical and coding tasks. It achieves comparable or better performance than many larger models while maintaining efficiency through its GQA attention mechanism.
Q: What are the recommended use cases?
The model excels in mathematical problem-solving, coding tasks, and complex reasoning scenarios. It's particularly well-suited for educational applications, automated problem-solving, and technical content generation where precise reasoning is required.
