EXAONE-3.5-32B-Instruct
| Property | Value |
|---|---|
| Parameters | 30.95B |
| Context Length | 32,768 tokens |
| Architecture | 64 layers, GQA with 40 Q-heads and 8 KV-heads |
| Vocabulary Size | 102,400 |
| License | EXAONE AI Model License Agreement 1.1 - NC |
| Author | LG AI Research |
What is EXAONE-3.5-32B-Instruct?
EXAONE-3.5-32B-Instruct is a state-of-the-art bilingual language model developed by LG AI Research, designed to excel in both English and Korean language tasks. As the flagship model in the EXAONE 3.5 series, it combines powerful performance with extensive context understanding capabilities of up to 32K tokens.
Implementation Details
The model employs a sophisticated architecture with 64 layers and uses Grouped-Query Attention (GQA) featuring 40 query heads and 8 key-value heads. It supports bfloat16 precision and can be deployed across various frameworks including TensorRT-LLM, vLLM, SGLang, llama.cpp, and Ollama.
- Advanced bilingual capabilities in English and Korean
- Optimized for instruction-tuning with system prompts
- Supports long-context processing up to 32K tokens
- Available in multiple quantized versions (AWQ and GGUF formats)
Core Capabilities
- Strong performance in MT-Bench (8.51) and LiveBench (43.0)
- Exceptional results in Arena-Hard (78.6) and IFEval (81.7)
- Superior Korean language understanding (KoMT-Bench: 8.05)
- Competitive performance against other leading models like Qwen 2.5 32B and Gemma 2 27B
Frequently Asked Questions
Q: What makes this model unique?
EXAONE-3.5-32B-Instruct stands out for its balanced bilingual capabilities and exceptional performance in both English and Korean. Its 32K token context window and optimized architecture make it particularly suitable for complex, long-form tasks.
Q: What are the recommended use cases?
The model excels in real-world applications requiring bilingual understanding, long-context processing, and complex reasoning. It's particularly effective for tasks requiring deep comprehension in either English or Korean, making it suitable for translation, content generation, and analytical tasks.
