EXAONE 3.5 32B
LG AI Research's flagship Korean-ecosystem model. Strong on Korean/Japanese language tasks; competitive on English. License blocks commercial use without LG agreement.
Positioning
EXAONE 3.5 32B is a dense 32-billion-parameter model released by LG AI Research under the EXAONE License. With a 32,768-token context window, it is designed primarily for Korean and Japanese language workloads, though it also performs competitively on English tasks. The model stands out in the open-weight landscape as a strong option for CJK (Chinese, Japanese, Korean) applications, but its license restricts commercial use without a separate agreement with LG.
Strengths
- Strong CJK language support: Built by LG AI Research, EXAONE 3.5 32B is specifically optimized for Korean and Japanese, making it a top choice for these languages.
- Dense architecture: As a dense 32B model, it avoids the complexity of mixture-of-experts (MoE) routing, simplifying deployment and inference.
- Generous context window: With 32,768 tokens of context, it can handle long documents, conversations, or code files without truncation.
- Multiple quantized variants: Quantized versions from Q8_0 down to Q2_K allow operators to trade off quality for memory footprint, with Q4_K_M at ~18 GB fitting many workstation GPUs.
Limitations
- Restrictive license: The EXAONE License prohibits commercial use without explicit agreement from LG AI Research, limiting deployment options for businesses.
- Large memory footprint at FP16: At ~64 GB in FP16, the model requires significant hardware; even quantized versions need careful memory planning.
- No community benchmarks available: We do not yet have independent, community-reported benchmarks for this model. Operators should treat published vendor metrics as best-case.
- English performance may lag behind dedicated English models: While competitive, the model's primary focus on Korean/Japanese means it may not match the English fluency of models trained predominantly on English data.
What it takes to run this locally
At FP16 precision, the model requires 64 GB of disk space and roughly 64 GB of VRAM, plus additional memory for KV cache and framework overhead (typically 30–50% more). This places it in the workstation deployment class: a single 48 GB GPU (e.g., RTX A6000) or dual 24 GB GPUs (e.g., two RTX 4090s) can run quantized versions. For example, Q4_K_M (18 GB) plus 9 GB for KV cache at half context fits within a single 24 GB card. Q2_K (10.4 GB) can run on a 12–16 GB consumer GPU, but with higher quality loss.
Should you run this locally?
Yes if: You need strong Korean or Japanese language support in a local, private deployment and can accept the license restrictions for your use case (e.g., research, personal projects). The dense architecture simplifies setup, and quantized versions make it accessible on workstation hardware.
No if: You require unrestricted commercial use, or your primary language is English and you need the absolute best English performance. In those cases, consider models with permissive licenses (e.g., Apache 2.0) or those specifically optimized for English.
Catalog cross-links
- EXAONE 3.0 7.8B
- EXAONE 3.5 2.4B
- LG AI Research
Overview
LG AI Research's flagship Korean-ecosystem model. Strong on Korean/Japanese language tasks; competitive on English. License blocks commercial use without LG agreement.
How to run it
EXAONE 3.5 32B is LG AI Research's 32B dense model. EXAONE is LG's model family optimized for Korean + English bilingual performance with competitive general reasoning. Run at Q4_K_M via Ollama (ollama pull exaone:32b) or llama.cpp with -ngl 999 -fa -c 8192. Q4_K_M file size ~18 GB on disk. Minimum VRAM: 16 GB — RTX 4080 (16GB) at Q4_K_M with KV offload. RTX 4090 24GB: Q4_K_M comfortably at 16K context. Recommended: RTX 4090 24GB at Q4_K_M. Throughput: ~35-55 tok/s on RTX 4090 at Q4_K_M. EXAONE architecture — compatible with standard inference stacks (llama.cpp support verified). EXAONE 3.5 is LG's most recent release — strong Korean performance, competitive English. Less benchmark coverage than Qwen/Mistral same-tier models, but quality is solid. Use for: Korean language tasks, bilingual (KO+EN) applications, general reasoning, coding. Not as strong on: non-Korean/English languages, niche domains. Context: 32K advertised; practical at Q4 on 24 GB is 16-32K. EXAONE is Apache 2.0 licensed — commercial-friendly.
Hardware guidance
Minimum: RTX 3060 12GB at Q3_K_M with KV offload. Recommended: RTX 4090 24GB at Q4_K_M (16K context). Optimal: RTX 5090 32GB at Q4_K_M (32K+ context). VRAM math: 32B dense, Q4_K_M ≈ 18 GB. KV cache at 16K: ~8 GB. Total: ~26 GB. RTX 4090 24GB: Q4 + 8-12K context fits on-GPU. 16K: offload KV. RTX 3090 24GB: same profile. RTX 4080 16GB: Q4 + 2K on-GPU. MacBook Pro M4 Pro 24GB+: Q4 at 10-20 tok/s. Cloud: A10 24GB at Q4_K_M. AWQ-INT4 drops to ~16 GB. EXAONE's 32B is hardware-efficient — one of the best Korean-capable models at consumer GPU size.
What breaks first
- Korean-centric tokenizer. EXAONE's tokenizer is optimized for Korean + English. Other languages (Japanese, Chinese, European) have higher token counts, reducing effective context. Test your language's token efficiency. 2. English quality vs Korean. English performance is competitive but lower than Korean. For English-only tasks, same-tier models like Qwen 3 32B may outperform. 3. LG licensing updates. LG may update EXAONE's license between versions. Verify Apache 2.0 status on the specific 3.5 release before commercial use. 4. Ecosystem maturity. EXAONE has less community quant coverage than Qwen/Mistral. Pre-converted GGUFs may be harder to find. Check bartowski/TheBloke for availability.
Runtime recommendation
Common beginner mistakes
Mistake: Assuming EXAONE matches Qwen/Mistral in English-only benchmarks. Fix: EXAONE is Korean-optimized. English quality is good but not best-in-class. Benchmark your English task against same-tier models before committing. Mistake: Using EXAONE for non-Korean/English languages. Fix: Tokenizer and training distribution are KO+EN heavy. Other languages underperform. Test your specific language. Mistake: Expecting broad GGUF availability. Fix: EXAONE has less community coverage. You may need to convert from hf yourself. Check bartowski's repo first. Mistake: Using Llama chat template. Fix: EXAONE uses LG's chat template. Verify on hf tokenizer_config.json. Using Llama template produces garbled Korean and awkward English.
Family & lineage
How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.
Strengths
- Best open Korean-language model in May 2026
- Strong CJK multilingual
Weaknesses
- License blocks unrestricted commercial use
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 |
|---|---|---|
| AWQ-INT4 | 19.0 GB | 22 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 EXAONE 3.5 32B.
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 EXAONE 3.5 32B?
Can I use EXAONE 3.5 32B commercially?
What's the context length of EXAONE 3.5 32B?
Source: huggingface.co/LGAI-EXAONE/EXAONE-3.5-32B-Instruct
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify EXAONE 3.5 32B runs on your specific hardware before committing money.