News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgEXAONE 4.5

EXAONE 4.5

Open-weight vision-language model from LG AI Research

LG AI Research released EXAONE 4.5, a 33-billion parameter multimodal model with open weights. It combines a causal language model with a visual encoder to handle text and images. The system supports a large context window and offers reasoning modes for complex tasks.

Share on WhatsApp

  • Total of 33 billion parameters including vision encoder
  • Supports context length of 262,144 tokens
  • Outperforms similar models in Korean reasoning tasks
  • Runs on single H200 GPU with optimized engines
  • First open-weight vision-language model from LG AI

What it is

EXAONE 4.5 is a vision-language model created by LG AI Research. It contains 33 billion parameters in total, with 1.2 billion dedicated to processing visual data. This model expands on the earlier EXAONE 4.0 framework by adding a visual encoder. It is designed to handle both text and image inputs effectively. The system is notable for its open-weight release, allowing developers to download and run it locally. It aims to provide strong performance in general benchmarks while excelling in specific areas like document analysis.

How it came about

This model builds directly upon the EXAONE 4.0 architecture. LG AI Research integrated a dedicated visual encoder into that existing framework to enable multimodal capabilities. The development focused on inheriting the powerful language skills of previous iterations while adding image understanding. The company released technical reports and code to support the launch. It represents the first open-weight vision-language model from this research team.

Inside the model

The architecture uses a causal language model paired with a vision encoder. It has 64 main layers and one multi-token prediction layer. The design employs grouped query attention and a hybrid pattern of sliding window and global attention. The context window extends to 262,144 tokens. Training data includes information up to December 2024. The vision part uses two-dimensional rotary positional embeddings.

What it can do

The model performs competitively on general benchmarks. It reportedly outperforms similar-sized state-of-the-art models in document understanding and Korean contextual reasoning. It supports agentic workflows, allowing it to use external tools like search engines. Users can switch between reasoning and non-reasoning modes. The reasoning mode is enabled by default for higher accuracy, while the other mode prioritizes speed. It prefers using a boxed format for final answers to improve parsing.

How to use it

LG AI Research provides the model weights openly. Developers can run it on various inference engines such as vLLM, TensorRT-LLM, and SGLang. It can operate on a single H200 GPU or four A100 GPUs with tensor parallelism. The setup creates an OpenAI-compatible API server for easy integration. Recommended settings vary by task, with specific temperature and penalty values suggested for OCR and general use.

Safety and testing

The developers note that the model may generate inappropriate or biased responses. These issues stem from statistical patterns in the training data. Efforts were made to remove harmful content, but some risks remain. The generated text does not reflect the views of LG AI Research. Users are warned about potential biases related to age, gender, and race.

Questions people ask about EXAONE 4.5

What hardware is needed to run EXAONE 4.5?

You can serve the model with its full 256K context length on a single H200 GPU. Alternatively, it runs on four A100-40GB GPUs using tensor parallelism. Optimized engines like vLLM or TensorRT-LLM are recommended for best performance.

Does EXAONE 4.5 support image inputs?

Yes, it is a vision-language model. It integrates a dedicated visual encoder with 1.29 billion parameters. This allows it to process images alongside text for tasks like document understanding and visual question answering.

How do I control the reasoning behavior?

Reasoning mode is enabled by default via the enable_thinking parameter. Set this to false if you need faster responses where latency matters more than accuracy. For high-accuracy tasks, keep it enabled.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on EXAONE 4.5 yet, so the facts on this page come from the sources listed above: huggingface.co/LGAI-EXAONE/EXAONE-4.5-33B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.