Notes

Uploaded distinct “thinking” and “non-thinking” versions of this 1.2B model. EXAONE 4.0’s blended local/global attention packs solid multilingual reasoning into a tiny footprint.

My recommendations:

• Thinking version: Use temperature=0.6 + top_p=0.95. Higher temps risk degeneration — add presence_penalty=1.5 if needed.
• Non-thinking version: Use temperature<0.6 (e.g., 0.1 for Korean to avoid code-switching). This is a Q8_0 quant.

Description

EXAONE 4.0 1.2B — LG AI’s compact multilingual model (EN/KO/ES) with distinct builds for fast chat and deeper problem-solving. Core innovation: Blended local/global attention for efficient 65k context handling in a tiny footprint.

Ideal for: - On-device multilingual Q&A and summarization - Lightweight math/coding with step-by-step reasoning (thinking version) - Agentic tool calling (e.g., function execution) - Educational/low-resource setups (Apache 2.0-like, but check EXAONE NC license for commercial use)

Enabling Reasoning Mode

The “thinking” version will provide step-by-step logic

References

EXAONE 4.0 on HuggingFace Technical Report GitHub LG AI Blog