EXAONE 3.5 2.4B

Positioning

EXAONE 3.5 2.4B is a dense 2.4B-parameter language model released by LG AI Research under the EXAONE AI Model License. It supports a 32,768-token context window and is designed for edge-tier deployment, with a particular strength in Korean-language chat. The model is bilingual (Korean/English) but carries a research-only license, limiting commercial use.

Strengths

• Compact footprint for edge deployment: At 2.4B parameters, quantized versions fit comfortably on consumer hardware. Q4_K_M (1.3 GB) and Q3_K_M (1.2 GB) allow the model to run on devices with limited memory, including phones or low-end GPUs.
• Long context window: With 32K tokens of context, the model can handle extended conversations or document processing without truncation, unusual for a model this small.
• Strong Korean language support: Purpose-built by LG AI Research for Korean chat, making it a top choice for Korean-language applications in the edge class.
• Low storage requirements: Even at FP16, the model is only ~5 GB; quantized versions drop below 2 GB, enabling easy distribution and fast loading.

Limitations

• Research-only license: The EXAONE AI Model License restricts use to non-commercial research. Operators seeking commercial deployment must look elsewhere.
• Small parameter count limits capability: As a 2.4B dense model, it cannot match the reasoning breadth or knowledge depth of larger models. It is best suited for narrow, well-defined tasks.
• No community benchmarks available: We do not have independent measurements of performance on standard tasks. Published vendor metrics should be treated as best-case.
• English performance may lag: While bilingual, the model's primary strength is Korean; English-language tasks may be better served by similarly sized models with stronger English training.

What it takes to run this locally

At FP16, the model requires ~5 GB of disk space. Quantized versions reduce this significantly: Q8_0 ~3 GB, Q6_K ~2.0 GB, Q5_K_M ~1.7 GB, Q4_K_M ~1.3 GB, Q3_K_M ~1.2 GB, Q2_K ~0.8 GB. Add 30–50% for KV cache and framework overhead at typical context lengths. The model fits in the edge deployment class — it can run on a single consumer GPU with 4–6 GB VRAM, or even on CPU with sufficient RAM.

Should you run this locally?

Yes if you need a compact, Korean-optimized chat model for research or prototyping, and you have limited hardware (e.g., a laptop or low-end GPU). The small quantized sizes make it easy to experiment with.

No if you require commercial use (license restriction), need broad English-language capability, or expect frontier-level reasoning. For production Korean chat, consider models with more permissive licenses.

Catalog cross-links

• EXAONE 3.5 7.8B
• EXAONE 3.5 32B
• Edge deployment guide