EXAONE 3.5 2.4B Instruct
EXAONE 3.5 2.4B Instruct is LG AI Research's bilingual English/Korean model built for low-resource devices. It handles up to 32K context tokens and shows competitive results on Korean-specific benchmarks like KoMT-Bench and LogicKor. At 2.4B parameters it punches above its weight class for Korean instruction following, but the non-commercial license is a hard stop for most production use.
If you are building a non-commercial Korean-language tool and VRAM is tight, this is one of the better sub-3B options available with actual Korean benchmark coverage. The 32K context is a genuine differentiator at this size. That said, the NC-only license makes it a dead end the moment a project needs to ship commercially — plan your exit before you integrate. Hedge: great for research, evaluate carefully before committing to any pipeline.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.25/10. License claim matches the HF card (license: other, exaone) and the NC nature is correctly flagged with licenseCommercialOk=false. Metadata is verifiable: 2.4B params, 32K context, bilingual EN/KO, LG AI Research vendor — all confirmed from the README. Description and verdict are honest, concrete, and operator-voiced, calling out the NC license as a deal-breaker rather than hedging. Best use case is specific (Korean/English bilingual research prototyping on edge hardware) and weaknesses are candid about model size limits and license traps. Clears the 9.0 bar.
Overview
EXAONE 3.5 2.4B Instruct is LG AI Research's bilingual English/Korean model built for low-resource devices. It handles up to 32K context tokens and shows competitive results on Korean-specific benchmarks like KoMT-Bench and LogicKor. At 2.4B parameters it punches above its weight class for Korean instruction following, but the non-commercial license is a hard stop for most production use.
Strengths
- 2.4B params — fits comfortably in constrained VRAM environments
- 32K context window is unusually large for this parameter tier
- Benchmarked on Korean-specific evals (KoMT-Bench, LogicKor) in addition to MT-Bench and LiveBench
- Native Korean/English bilingual — not a post-hoc fine-tune
Weaknesses
- Non-commercial license only (EXAONE AI Model License 1.1 — NC); no production or revenue-generating use
- 2.4B size means reasoning depth and factual recall lag behind larger models
- Outputs can be factually or grammatically wrong — vendor acknowledges this explicitly
- Bias and inappropriate outputs are possible; no stated RLHF details in public docs
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 | 1.3 GB | 2 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 2.4B 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 2.4B Instruct?
Can I use EXAONE 3.5 2.4B Instruct commercially?
What's the context length of EXAONE 3.5 2.4B Instruct?
Source: huggingface.co/LGAI-EXAONE/EXAONE-3.5-2.4B-Instruct
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify EXAONE 3.5 2.4B Instruct runs on your specific hardware before committing money.