LG EXAONE
Korea's open vision-language AI — EXAONE 4.5 beats GPT-5-mini on STEM
Overview
LG AI Research's EXAONE family added vision to its hybrid reasoning models with EXAONE 4.5 (April 2026): a 33B open-weight VLM (32B LLM + 1.2B vision encoder) that reads text and images, now speaks six languages including Japanese and Vietnamese, and averages 77.3 across five STEM benchmarks.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- EXAONE AI Model License (non-commercial)
- Multiple sizes incl. reasoning variants
- Self-host on your own hardware
- K-EXAONE 236B: $0.20/M in, $0.80/M out
- Cached input $0.10/M tokens
- OpenAI-compatible endpoints
- Contact LG AI Research
- Enterprise deployment support
Use cases
What you can produce with LG EXAONE
- Self-hostable 32B hybrid reasoning model weights (BF16, GPTQ, AWQ, GGUF) via Hugging Face
- Compact 1.2B on-device model optimised for edge and mobile deployment
- 128K-token context window supporting long-document and multi-turn agentic tasks
- Structured tool-calling / function-calling support (BFCL-v3: 63.9%) for agent pipelines
- Commercial API access via FriendliAI Serverless and Dedicated Endpoints (pay-per-token)
- Korean, English, and Spanish multilingual output with strong KMMLU-Pro performance (67.7%)
- Compatibility with vLLM, TensorRT-LLM, llama.cpp, and HuggingFace Transformers v4.54.0+
ASEAN Perspective
LG EXAONE in Southeast Asia
EXAONE 4.0 does not currently support any Southeast Asian language — Vietnamese, Thai, Indonesian, Malay, Filipino, and other regional languages are absent — making it a weaker out-of-the-box choice for ASEAN-language NLP tasks compared to Alibaba Qwen or SeaLLMs. That said, its strong math, science, and coding benchmarks are language-agnostic, and ASEAN enterprises running Korean-joint-venture or Korean-tech supply-chain operations may find the model's Korean-English bilingual strength directly useful. LG has a substantial hardware and electronics presence across Southeast Asia (consumer electronics, home appliances, automotive), and as the EXAONE ecosystem matures — particularly the on-device 1.2B variant for embedded applications — ASEAN device manufacturers partnering with LG could encounter EXAONE in embedded product contexts before any cloud deployment arrives.
EXAONE 4.5, out in April 2026, moves LG AI Research's hybrid reasoning line into vision: a 33B model (32B language backbone plus a 1.2B vision encoder) that reads images alongside text and averages 77.3 across five STEM benchmarks — ahead of GPT-5-mini (73.5), Claude Sonnet 4.5 (74.6), and roughly level with Qwen3 235B (77.0), despite a fraction of that last model's parameter count. Language support has also widened: EXAONE 4.5 now covers Korean, English, Spanish, German, Japanese, and Vietnamese, putting at least one ASEAN language on the table for the first time in this family. The friction points from EXAONE 4.0 mostly carry over. Weights ship under the non-commercial EXAONE AI Model License 1.2-NC; production use needs a separate deal, typically through inference partner FriendliAI. Developer tooling and third-party integrations remain thinner than for Qwen or Llama releases, and LG's own product lines (appliances, laptops, robotics) stay the primary use case driving development.
What people say
77.3 versus 73.5: that's EXAONE 4.5 beating OpenAI's GPT-5-mini on an average of five STEM benchmarks, and it also edges past Claude Sonnet 4.5 (74.6) and roughly ties Qwen3 235B (77.0) — a model seven times its size. LG AI Research shipped EXAONE 4.5 on 9 April 2026, nine months after 4.0, and the headline change is vision: a 1.2B-parameter encoder bolted onto the existing 32B reasoning backbone, so the model now reads charts, documents, and images instead of text alone.
The language list grew too, and this is the detail worth flagging for anyone reading from Southeast Asia: EXAONE 4.5 adds Japanese and Vietnamese to its earlier Korean/English/Spanish set. That's a real change from the 4.0 generation, which had zero ASEAN coverage. It's still a narrow list next to Qwen or SEA-LION, but Vietnamese support is new and worth noting rather than writing off.
What hasn't changed is the licensing model. EXAONE ships under a non-commercial research license (EXAONE AI Model License 1.2-NC); anyone wanting to run it in production needs a separate commercial agreement, in practice arranged through LG's inference partner FriendliAI. Community tooling is still thin compared to fully open licenses like Apache 2.0 — there's an official GitHub org with reference code, but the plugin and integration ecosystem that's grown up around Qwen or Llama doesn't really exist here yet. Best read as a serious, benchmark-backed research release from a company that mainly wants it running inside its own appliances and robots, not as a drop-in open API.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- EXAONE 4.0's 32B model scored 94.5 on South Korea's university entrance exam (CSAT) math section and 90 on the 2024 US Mathematical Olympiad.
- The name EXAONE is an acronym for 'EXpert AI for EveryONE', reflecting LG's goal of broad enterprise deployment rather than a single consumer product.
- EXAONE 4.0's 1.2B on-device model is half the size of its predecessor (EXAONE 3.5 2.4B) yet outperforms OpenAI's GPT-4o mini on specialist math and coding evaluations.
- LG AI Research's founding president Kyunghoon Bae was nominated as South Korea's Minister of Science and ICT in 2025, underscoring the model's national-strategic significance.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including LG EXAONE's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with LG EXAONE unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to LG EXAONE directly →
Spotted something out of date? Suggest an update →
Alternatives to LG EXAONE
More in LLMs & Chat
