Hancheol Park

AI Research Engineer

Insight

Ternary quantization results: model size drops from 60GB in BF16 to about 10GB at 1.58 bits (about 1/6), with performance at 89% of BF16

November 10, 2026

min read

Can a 30B Local LLM Run on a MacBook? 1.58-Bit Ternary Quantization

Graphic showing GPU reduction for Solar Open2 and Kimi K3. A stack of 8 GPUs shrinks to 2 for Solar Open2 and to 4 for Kimi K3.

August 26, 2026

min read

From Eight H100s to Two: The GPU Cost Savings of LLM Compression

min read

Two papers discussing quantization methods specialized for MoE LLMs, accepted to an ICML 2026 workshop

min read

Deploying an Efficient Vision-Language Model on Mobile Devices

min read

NotaMoEQuantization: An MoE-Specific Quantization Method for Solar-Open-100B

February 10, 2025

min read

Where do LLMs Encode the Knowledge to Assess the Ambiguity?

Culture

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.