EXAONE Deep offers a suite of reasoning-enhanced Large Language Models (LLMs) ranging from 2.4B to 32B parameters, designed for advanced mathematical and coding tasks. It targets researchers and developers seeking high-performance models for complex problem-solving, outperforming comparable open-weight models and even proprietary solutions in specific benchmarks.

How It Works

EXAONE Deep models are built on a proprietary architecture optimized for reasoning. The models leverage specific prompting strategies, including the use of \n tokens to delineate reasoning steps and a recommended prompt structure for math and coding problems (e.g., "Please reason step by step, and put your final answer within \boxed{}."). This approach aims to improve the coherence and accuracy of step-by-step problem-solving.

Quick Start & Requirements

• Install transformers>=4.43.1.
• Models can be loaded directly via Hugging Face Transformers: AutoModelForCausalLM.from_pretrained("LGAI-EXAONE/EXAONE-Deep-7.8B", torch_dtype=torch.bfloat16, device_map="auto").
• Quantized versions (AWQ, GGUF) are available for local execution via llama.cpp, Ollama, and LM-Studio.
• Full documentation and examples are available on Hugging Face and linked within the README.

Highlighted Details

• Outperforms comparable models in math and coding benchmarks (e.g., MATH-500, AIME, Live Code Bench).
• 7.8B model surpasses OpenAI's o1-mini on reasoning tasks.
• Supports deployment via TensorRT-LLM, vLLM, and SGLang.
• Offers AWQ and GGUF quantized versions for efficient inference.

Maintenance & Community

Developed and released by LG AI Research. Contact available via contact_us@lgresearch.ai.

Licensing & Compatibility

Licensed under the EXAONE AI Model License Agreement 1.1 - NC (Non-Commercial). This license restricts commercial use.

Limitations & Caveats

The model may generate inappropriate, biased, or factually incorrect responses due to its reliance on training data statistics. It does not reflect the latest information and users are prohibited from inducing inappropriate outputs that violate LG AI's ethical principles.