GLM-5.3 open weights released; Unsloth enables local run with 83% size reduction
danielhanchen · x · 2026-08-29
Z.ai open-sourced GLM-5.3, its most capable model (744B params, 40B active), achieving SOTA on Terminal Bench 3.0. Unsloth released a local guide, drastically lowering hardware barriers via dynamic quantization:
- 2-bit Quantization: Reduces model size from 1.51TB to 239GB (-83%) while retaining 81% accuracy.
- Hardware Requirements: The 2-bit version runs on a Mac with 256GB RAM or mixed RAM/VRAM setups (recommended 1x24GB GPU + 256GB RAM).
- Usage: Unsloth Desktop allows toggling between Low/High/Max thinking modes via UI; Max Thinking is recommended for complex coding tasks.
More from coding & agent
- Building a Self-Maintaining Knowledge System for Claude — PawelHuryn · 2026-08-29
- Claude Tag: A Promptable Harness and Software Factory — maxleiter · 2026-08-29
- Agent-Agent Communication Overcomes Single-Machine Compute Limits — peterjliu · 2026-08-29
- Google Paper Reveals Agent Skill Compounding Mechanism — PawelHuryn · 2026-08-29
- Claude Code 2.1.251 Adds Model Switch Hooks — ClaudeCodeLog · 2026-08-29
- Claude Code 2.1.251 Released with Model Switch Hooks and Security Limits — ClaudeCodeLog · 2026-08-29