LG AI Research · 2026-04-09 · notable
EXAONE 4.5 — LG's first open-weight vision-language model
LG AI Research releases EXAONE 4.5, a 33B vision-language model fusing the EXAONE 4.0 hybrid LLM with a 1.29B vision encoder. 262K context, six languages, with day-one TensorRT-LLM, vLLM, SGLang and llama.cpp support.
LG's EXAONE family finally goes multimodal — a 33B VLM with 262K context, document-understanding focus, and full inference-stack support on day one.
Key specs
| License | EXAONE AI Model License 1.2 (NC) |
|---|---|
| Context window | 262,144 tokens |
| Total parameters | 33B |
| Lm parameters | 31.7B |
| Vision encoder | 1.29B |
| Stem visual reasoning (lg internal) | 77.3 |
What is it?
EXAONE 4.5 is the first open-weight vision-language model from LG AI Research, the in-house AI lab inside LG. It bolts a 1.29B-parameter vision encoder onto the existing EXAONE 4.0 hybrid LLM (31.7B parameters), giving the model image-text-to-text capability while keeping the dual reasoning / non-reasoning mode design that EXAONE 4.0 introduced. The technical report was posted to arXiv on April 9, 2026.
How does it work?
The language backbone is the EXAONE 4.0 hybrid architecture, which uses a 16-layer repeating block of 3 sliding-window attention layers followed by 1 global attention layer to support 262K-token context efficiently. The vision encoder feeds image tokens into the same shared transformer. Training emphasized curated document-understanding data, and the model is released for research/academic/educational use under the EXAONE AI Model License 1.2 (non-commercial). LG ships day-one support for TensorRT-LLM, vLLM, SGLang, and llama.cpp, plus official GGUF quantizations.
Why does it matter?
Most open-weight VLMs come from a small set of labs (Qwen, InternLM, Molmo, Gemma). LG joining that group with a 33B model focused on document understanding and Korean / multilingual reasoning is a meaningful addition — especially for non-English enterprise document workflows. LG's internal benchmarks claim a 77.3 average on five visual STEM reasoning metrics, ahead of GPT-5 mini (73.5) and Claude Sonnet 4.5 (74.6). The model card has already crossed 4,000 downloads in its first week.
Who is it for?
Researchers and self-hosters working on document AI, Korean-language NLP, or multilingual VLM evaluation.
Try it
huggingface.co/LGAI-EXAONE/EXAONE-4.5-33B