EXAONE 3.5 7.8B Instruct
EXAONE 3.5 7.8B is LG AI Research's instruction-tuned bilingual model for English and Korean, with a 32K token context window. It succeeds EXAONE 3.0 with reported benchmark gains on MT-Bench and LiveBench. The non-commercial license is a hard blocker for production use.
If you need a capable Korean-English model for personal experimentation or research, EXAONE 3.5 7.8B is a solid pick at this parameter count. Benchmark numbers are competitive and the 32K context is genuinely useful. That said, the NC-only license is a hard stop — if there's any chance your use case touches commercial ground, look elsewhere. Hedge: worth a local test run, not worth building on.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.05/10. Metadata is fully verifiable against the model card: 7.8B params, 32K context, bilingual EN/KO, LG AI Research, license:other (EXAONE NC). The row correctly flags the non-commercial license as a hard blocker, which is the single most important fact for a runlocalai reader. Description is concrete and operator-voiced, weaknesses are honest, and the verdict gives a clear 'test locally, don't ship' recommendation. Minor nits: bestUseCase 'Korean-English research and personal projects' is slightly broad — could be sharper (e.g., 'Korean-English bilingual chat for local prototyping'), and family='other' is acceptable but EXAONE is its own family worth noting. Overall a clean, publishable row.
Flags: - bestUseCase slightly generic — 'research and personal projects' could be tighter - License version stated as '1.1' — model card just says 'exaone' with LICENSE link; verify the 1.1 version number is accurate
Overview
EXAONE 3.5 7.8B is LG AI Research's instruction-tuned bilingual model for English and Korean, with a 32K token context window. It succeeds EXAONE 3.0 with reported benchmark gains on MT-Bench and LiveBench. The non-commercial license is a hard blocker for production use.
Strengths
- 32K context window handles long documents and extended conversations
- Bilingual English/Korean — one of the stronger Korean-capable models at this size
- Runs on common local stacks: llama.cpp, Ollama, vLLM, TensorRT-LLM
- 205K+ HF downloads suggests reasonable community testing and feedback
Weaknesses
- Non-commercial license only — cannot be used in any product or paid service
- Knowledge cutoff means factual answers can be stale or wrong
- Known risk of biased or inappropriate outputs per vendor's own disclosure
- May produce fluent but semantically incorrect sentences — verify critical outputs
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 | 4.3 GB | 6 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 7.8B Instruct.
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 7.8B Instruct?
Can I use EXAONE 3.5 7.8B Instruct commercially?
What's the context length of EXAONE 3.5 7.8B Instruct?
Source: huggingface.co/LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify EXAONE 3.5 7.8B Instruct runs on your specific hardware before committing money.