K-EXAONE 236B A23B
K-EXAONE is LG AI Research's 236B Mixture-of-Experts model with 23B active parameters per forward pass. It covers Korean, English, Spanish, German, Japanese, and Vietnamese with a 262K token context window. Benchmarks show solid math and coding performance, but the custom license blocks commercial use.
If you need a strong Korean-first model for research or internal non-commercial work and have the hardware to run it, K-EXAONE is worth a look — the MoE design makes 236B more tractable than it sounds. That said, the custom license is a hard blocker for any commercial deployment, full stop. At 55K downloads it has real traction, but the lack of wide independent benchmarking means you should test it on your own workloads before committing infrastructure. Hedge: promising for research shops, skip if you need production commercial rights.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.02/10. License claim matches the HF card (license: other, license_name: k-exaone) and commercialOk=false is the safe correct default for a custom non-OSI license. Metadata is accurate: 236B total / 23B active MoE, LG AI Research vendor, multilingual coverage all verified. Note the card says native 256K context but the row uses 262144 (262K) — technically 256K = 262144 tokens, so this is correct. Editorial voice is honest, flags vendor-reported benchmarks and license traps, and the verdict gives a clear hedge. Use case 'Korean-language research and non-commercial reasoning' is specific enough given the model's actual positioning. Deployability section honestly addresses the multi-GPU reality even with MoE.
Flags: - Description says '262K token context window' — accurate numerically (262144) but the vendor markets it as 256K; minor wording inconsistency worth a copy pass - arxiv:2601.01739 in tags is suspicious (future-dated arxiv ID format) but this is an HF tag issue, not the row's fault
Overview
K-EXAONE is LG AI Research's 236B Mixture-of-Experts model with 23B active parameters per forward pass. It covers Korean, English, Spanish, German, Japanese, and Vietnamese with a 262K token context window. Benchmarks show solid math and coding performance, but the custom license blocks commercial use.
Strengths
- MoE architecture: 236B total params, only ~23B active at inference — lowers VRAM pressure versus dense 236B models
- 262K token context window via sliding window attention
- Multilingual 150k-vocabulary SuperBPE covering 6 languages including Korean
- Competitive math and coding benchmark scores per vendor reporting
Weaknesses
- Custom 'k-exaone' license — commercial use is not permitted
- Even with MoE efficiency, serving a 236B model demands serious multi-GPU hardware
- Vendor-reported benchmarks; independent third-party evals are limited at time of listing
- Korean is the primary optimization target; other language quality may vary
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 129.8 GB | 166 GB |
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of K-EXAONE 236B A23B.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run K-EXAONE 236B A23B?
Can I use K-EXAONE 236B A23B commercially?
What's the context length of K-EXAONE 236B A23B?
Source: huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify K-EXAONE 236B A23B runs on your specific hardware before committing money.