EXAONE 4.0
LG's open-weight hybrid LLM-plus-reasoning model: a 32B expert and a 1.2B on-device model, Korea's first hybrid AI.
About
EXAONE 4.0 is LG AI Research's flagship model, unveiled July 15, 2025 as Korea's first open-weight hybrid AI — an LLM for rapid knowledge responses fused with a reasoning model that forms and verifies hypotheses, following the country's first reasoning model (EXAONE Deep, March 2025). It ships in two sizes: a 32B expert model that LG says passed written exams for six Korean professional licenses (physicians, dentists, customs brokers, appraisers and more), and a 1.2B on-device model for phones, appliances, cars and robots that LG claims beats GPT-4o mini on math, coding and science evals despite being half the size of its predecessor.
On LG's own benchmark table the 32B scores 92.3 on MMLU-Redux, 81.8 on MMLU-Pro, 66.7 on LiveCodeBench v6, 75.4 on GPQA-Diamond and 85.3 on AIME 2025; independent Artificial Analysis testing puts MMLU-Pro at 76.8% and GPQA at 62.8% — solid, if below the vendor's framing. Weights are public on Hugging Face but under a custom research/academic/educational license, not an OSI-approved one, with a commercial API offered through FriendliAI. LG followed it with K-Exaone, which ranked 7th on the Artificial Analysis Intelligence Index in early 2026 — the only Korean model in the global top 10.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- MMLU-Redux 92.3, MMLU-Pro 81.8, LiveCodeBench v6 66.7, GPQA-Diamond 75.4, AIME 2025 85.3 (LG-reported)
- MMLU-Pro 76.8%, GPQA 62.8%, LiveCodeBench 47.2%, MATH-500 93.9%, AIME 2025 39.3% (third-party, Artificial Analysis via dataconomy, Apr 2026)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: LG AI Research release Jul 15, 2025 with open weights on HF; LG-reported benchmark table plus independent Artificial Analysis replication (MMLU-Pro 76.8%); follow-up K-Exaone ranked 7th on AA Intelligence Index Jan 2026. Open weights. Release facts corroborated across Korea JoongAng Daily, The Investor, and TNGlobal. Strengths: genuine hybrid reasoning design, rare Korean-first frontier effort, third-party evals exist. Docked for: (1) headline benchmarks are vendor-run with the AA replication notably lower; (2) custom research-only license, not OSI-open, limiting commercial reuse of the weights.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · LG AI Research release 2025-07-15; 32B expert + 1.2B on-device; open weights on HF; K-Exaone 7th on AA Intelligence Index Jan 2026
- · Docked: Headline benchmarks vendor-run; AA replication notably lower (MMLU-Pro 76.8 vs 81.8 claimed)
- · Docked: Research-only license, not OSI-open; limits commercial reuse