LG AI Research shipped EXAONE 4.5 on June 12, 2026, a 32 billion parameter open-weight vision language model that can process images and text together. The weights are available through Hugging Face and the official LG AI Research announcement page. The release includes safetensor checkpoints, tokenizer files, and an inference guide. EXAONE 4.5 supports a 128,000-token context window and can handle documents, charts, handwritten notes, and natural images. It is a direct answer to teams that need capable vision language reasoning without paying per-token closed API rates. This launch matters for free AI tool users and builders who track open-weight model releases.
EXAONE 4.5 lands at a moment when many commercial AI providers have tightened free API tiers and pushed flagship models behind paid plans. The model’s open license removes per-image or per-token charges for self-hosted use. LG AI Research reports that EXAONE 4.5 scores 92.8 on DocVQA, 88.1 on ChartQA, and 61.7 on MMMU. Those numbers place it above several closed mid-tier vision models on document and chart reasoning. For developers and IT leaders watching AI price changes in June 2026, EXAONE 4.5 offers a path to fixed-cost inference.
The model is a dense transformer with a vision encoder that maps image patches into the same embedding space as text tokens. It was trained on multilingual text and image instruction data, with extra emphasis on Korean, English, Japanese, and document understanding tasks. LG AI Research published benchmark comparisons against OpenAI GPT-4o, Google Gemini 2.0 Flash, and Qwen2.5-VL. On the MMMU multimodal reasoning benchmark, EXAONE 4.5 trails GPT-4o by less than one point while beating Gemini 2.0 Flash on ChartQA and DocVQA. The context window of 128,000 tokens means users can upload long PDFs and multi-image sequences in a single prompt. This technical profile is relevant for teams following major model tier changes.
EXAONE 4.5 is released under an EXAONE AI Open License that permits commercial use, modification, and redistribution with an attribution notice. The license does not require a paid subscription or API key for self-hosting. That separates the release from closed APIs that have introduced stricter free tier limits in 2026. The model requires about 19 GB of memory in 4-bit quantization and can run on a single 24 GB GPU. This makes it plausible for local inference on workstation hardware.
How Do the Top Options Compare?
| Model | Best For | Parameters | Context | License | Vision Benchmarks |
|---|---|---|---|---|---|
| EXAONE 4.5 | Open document and chart reasoning | 32B | 128K tokens | EXAONE AI Open License | DocVQA 92.8, ChartQA 88.1, MMMU 61.7 |
| Qwen2.5-VL 72B | High-capacity open vision tasks | 72B | 128K tokens | Apache 2.0 | DocVQA 93.1, MMMU 64.5 |
| Llama 3.2 Vision 11B | Lightweight local image tasks | 11B | 128K tokens | Llama 3.2 Community License | DocVQA 86.0, MMMU 50.2 |
| Gemini 2.0 Flash | Closed API multimodal automation | Unknown | 1M tokens | Proprietary API | DocVQA 90.4, ChartQA 85.6, MMMU 63.8 |
| MiniMax-VL 7B | Small edge vision fine-tunes | 7B | 32K tokens | Open license | DocVQA 82.0, ChartQA 74.5 |
Benchmarks are vendor-reported and may differ across evaluation pipelines. EXAONE 4.5 scores are from LG AI Research. Closed model scores use public benchmark reports. Local hardware requirements depend on quantization and batch size.
1. EXAONE 4.5 , Open-weight document and chart reasoning
EXAONE 4.5 is LG AI Research’s open-weight vision language model. It has 32 billion parameters, a 128,000-token context window, and native support for image plus text inputs. The model is tuned for DocVQA, ChartQA, and real-world user interfaces. It is available from LG AI Research and Hugging Face as safetensor weights with a tokenizer and example scripts. This release gives self-hosted teams a model that can read dense documents, extract table data, and answer questions about charts without calling a closed API.
On the evaluation side, LG AI Research reports DocVQA at 92.8, ChartQA at 88.1, and MMMU at 61.7. Those scores beat Gemini 2.0 Flash on document and chart tasks and come close to GPT-4o on broader multimodal reasoning. The model also handles Korean and Japanese text with stronger performance than previous EXAONE releases. For developers who have watched AI pricing changes in June 2026, the model removes variable per-image costs.
Local deployment is practical. In 4-bit quantization, EXAONE 4.5 uses about 19 GB of GPU memory, which fits a single RTX 4090 or A10. It can also run quantized on CPU with slower but workable speeds for small batch jobs. The model does not yet have native video input, and code generation from visual diagrams is less mature than text coding models. These limits are honest tradeoffs for the license freedom.
Key strengths:
- ✅ Open license allows commercial self-hosting without per-token fees
- ✅ Strong DocVQA and ChartQA performance for document workflows
- ✅ 128K context handles long PDFs and multiple images
- ✅ Runs on a single 24 GB GPU in 4-bit quantization
- ❌ No native video understanding in the initial release
- ❌ Visual diagram to code conversion is not the model’s strength
- ❌ Requires real GPU memory for full precision inference
Who it’s for: Choose EXAONE 4.5 if you need an open vision language model for document parsing, chart QA, or self-hosted multimodal agents.
2. Qwen2.5-VL 72B , High-capacity open vision reasoning
Qwen2.5-VL 72B is Alibaba’s open vision language model with strong high-resolution image understanding. It has 72 billion parameters and a 128,000-token context. The model is available under the Apache 2.0 license, which gives firms broad modification and redistribution rights. Qwen2.5-VL supports images, video, and text, making it a flexible choice for multimodal pipelines. It often leads open model leaderboards on MMMU and document benchmarks.
For document use, Qwen2.5-VL reports DocVQA near 93.1 and MMMU around 64.5. Those numbers are slightly ahead of EXAONE 4.5 on MMMU but require more memory and compute. The larger parameter count means a single GPU is harder to use unless teams deploy tensor parallel across multiple cards or use aggressive 3-bit quantization. The model is available through Hugging Face and Qwen’s official repositories.
The model fits organizations that already have multi-GPU infrastructure. It also includes video input, which EXAONE 4.5 lacks in the first release. However, Apache 2.0 does not solve the compute cost problem. Teams that need frequent document parsing may find EXAONE 4.5 more efficient despite slightly lower MMMU. For tracking open source releases, see major open model changes.
Key strengths:
- ✅ Apache 2.0 license with wide commercial freedom
- ✅ Higher MMMU score than EXAONE 4.5
- ✅ Native video and high resolution image support
- ❌ 72B parameters demand multi-GPU or very low quantization for local use
- ❌ Higher memory and latency costs than smaller open models
- ❌ Deployment complexity is greater for small teams
Who it’s for: Choose Qwen2.5-VL 72B if you need maximum open vision capability and already run multi-GPU infrastructure.
3. Llama 3.2 Vision 11B , Lightweight local image tasks
Meta’s Llama 3.2 Vision 11B is a compact open vision language model for lightweight local image understanding. It has 11 billion parameters, supports a 128K context window, and runs well on a single 16 GB GPU. The model is tuned for image captioning, object identification, and simple visual QA. It is not a document chart specialist, but its small size makes it easy to deploy in edge and mobile-adjacent environments.
The Llama 3.2 Community License is not fully open by Open Source Initiative standards. It includes restrictions for services with very large monthly active users. Still, the model can be self-hosted and modified for most small and mid-size deployments. Reported benchmarks put DocVQA around 86.0 and MMMU near 50.2. Those scores trail EXAONE 4.5 and Qwen2.5-VL, but the footprint is far lighter.
Developers who need a simple image understanding model for an on-device app or an internal tool may prefer Llama 3.2 Vision. The smaller size also means faster cold starts and lower memory use. But if the task is reading dense forms, invoices, or complex charts, EXAONE 4.5 is the stronger open option. This comparison matters as AI free tier limits get tougher.
Key strengths:
- ✅ Small 11B footprint runs on a single 16 GB GPU
- ✅ Simple image captioning and visual QA are fast
- ✅ Broad ecosystem support in frameworks and tooling
- ❌ Document and chart accuracy is lower than EXAONE 4.5
- ❌ License includes use restrictions for very large deployments
- ❌ Struggles on complex multi-step visual reasoning
Who it’s for: Choose Llama 3.2 Vision 11B if you need a lightweight, local image model for simple tasks and edge use.
4. Gemini 2.0 Flash , Closed API multimodal automation
Google AI’s Gemini 2.0 Flash is a closed API model for fast, large-scale multimodal automation. It offers a 1 million token context window and strong vision, video, and text understanding. Developers access it through an API key, with per-token and per-image costs. Google has been adjusting Gemini free tier access and pricing, which makes long-term cost harder to predict.
On the same benchmark set, Gemini 2.0 Flash reports DocVQA around 90.4, ChartQA near 85.6, and MMMU about 63.8. Its MMMU score beats EXAONE 4.5, while its chart score comes in lower. The huge context window is useful for very long video or document inputs. But the model is not open-weight, so teams cannot self-host or inspect the weights. That creates vendor dependency.
For rapid prototyping and agent workflows, Gemini 2.0 Flash is convenient. It requires no GPU infrastructure and has excellent integration with Google Cloud and Vertex AI. However, ongoing AI pricing changes and API quotas can surprise teams with variable bills. EXAONE 4.5 offers a fixed-cost self-hosted alternative for stable internal workloads.
Key strengths:
- ✅ Very large 1 million token context window
- ✅ Strong MMMU score and video understanding
- ✅ No local GPU setup required
- ❌ Closed weights prevent self-hosting and inspection
- ❌ Per-token and per-image pricing can become unpredictable
- ❌ Free API tier limits have tightened in June 2026
Who it’s for: Choose Gemini 2.0 Flash if you need a no-ops closed API for rapid multimodal automation and have budget for variable costs.
5. MiniMax-VL 7B , Small edge vision fine-tunes
MiniMax’s M-series open-weight models include a compact vision language option for lightweight multimodal tasks. MiniMax-VL 7B offers 7 billion parameters, a 32,000-token context window, and an open license. It runs on smaller server GPUs and is easier to fine-tune than 32B models. The model is available through MiniMax and Hugging Face.
It is not a document chart specialist. MiniMax-VL reports lower DocVQA and ChartQA scores than EXAONE 4.5, but its size makes it useful for quick image captioning and sorting tasks. The shorter context window also means it cannot process very long PDFs in a single call. Teams that need a smaller patch for their own fine-tuning may prefer this model.
For developers watching open-weight model releases, MiniMax-VL is another sign that small vision models are getting closer to enterprise use. But if the core workload is document-heavy, EXAONE 4.5’s stronger DocVQA and larger context are better tradeoffs.
Key strengths:
- ✅ Small 7B model is easy to fine-tune
- ✅ Low memory footprint fits edge server GPUs
- ✅ Open license with commercial use
- ❌ Weaker DocVQA and ChartQA than EXAONE 4.5
- ❌ 32K context is short for long documents
- ❌ Limited support for complex visual reasoning
Who it’s for: Choose MiniMax-VL 7B if you want a small open vision model for edge deployment and custom fine-tuning.
Frequently Asked Questions
What is EXAONE 4.5?
EXAONE 4.5 is a 32 billion parameter open-weight vision language model from LG AI Research. It processes images and text together and supports a 128,000-token context window. The model is built for document, chart, and visual reasoning tasks.
When was EXAONE 4.5 released and where can I get it?
EXAONE 4.5 was released on June 12, 2026. The weights are available on Hugging Face and on the LG AI Research homepage. You can download safetensor checkpoints, tokenizer files, and inference examples without an API key.
What license does EXAONE 4.5 use?
EXAONE 4.5 is released under the EXAONE AI Open License. The license allows commercial use, modification, and redistribution with an attribution notice. It supports self-hosting without per-token or per-image fees.
How does EXAONE 4.5 compare to closed models like GPT-4o and Gemini 2.0 Flash?
LG AI Research reports that EXAONE 4.5 scores 92.8 on DocVQA and 88.1 on ChartQA, beating Gemini 2.0 Flash on those tasks. On MMMU it scores 61.7, which is close to GPT-4o and slightly behind Gemini 2.0 Flash. The tradeoff is open weight self-hosting versus closed API convenience.
What hardware do I need to run EXAONE 4.5?
In 4-bit quantization, EXAONE 4.5 needs about 19 GB of GPU memory. A single 24 GB GPU like an RTX 4090 or A10 can run the quantized model. Full precision inference requires more VRAM and is better suited to multi-GPU systems.
Does EXAONE 4.5 support video input?
The initial EXAONE 4.5 release does not include native video understanding. It supports text and image inputs, including documents, charts, and multi-image sequences. Teams that need video may look at Qwen2.5-VL or Gemini 2.0 Flash.
What Should You Remember?
- Open-weight release: EXAONE 4.5 from LG AI Research is a 32B vision language model available for self-hosting.
- Document strength: It scores 92.8 on DocVQA and 88.1 on ChartQA, beating Gemini 2.0 Flash on chart tasks.
- License freedom: The EXAONE AI Open License permits commercial use without per-token or per-image API fees.
- Hardware fit: 4-bit quantization runs on a single 24 GB GPU with about 19 GB of memory.
- Closed model tradeoff: Gemini 2.0 Flash has a larger context and higher MMMU but cannot be self-hosted.
- Limitations: The first release lacks native video input and is not a visual diagram to code specialist.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
