GLM-5.3-Flash
Z.ai's MIT-licensed efficient model — natively multimodal, with hybrid sparse and linear attention far cheaper to run than GLM-5.3
GLM-5.3-Flash
Z.ai • August 2026
Training Data
Not disclosed
GLM-5.3-Flash
August 2026
Parameters
320 billion (18B active)
Training Method
MoE with hybrid sparse and linear attention
Context Window
1,000,000 tokens
Knowledge Cutoff
Not disclosed
Key Features
1M Context / 128K Output • Text, Image & Video Input • MIT-Licensed Open Weights • Hybrid Sparse + Linear Attention
Capabilities
Coding: Excellent (63.4 DeepSWE v1.1)
Automation: Excellent (48.8 AutomationBench)
Cost Efficiency: Outstanding
What's New in This Version
Beats GLM-5.2 on DeepSWE v1.1 (63.4 vs 46.2) and AutomationBench (48.8 vs 26.2) with 18B active parameters, cutting attention compute 3.01x and KV cache 4.44x versus GLM-5.3; $0.15/$0.50 per 1M tokens at list price
Z.ai's MIT-licensed efficient model — natively multimodal, with hybrid sparse and linear attention far cheaper to run than GLM-5.3
What's New in This Version
Beats GLM-5.2 on DeepSWE v1.1 (63.4 vs 46.2) and AutomationBench (48.8 vs 26.2) with 18B active parameters, cutting attention compute 3.01x and KV cache 4.44x versus GLM-5.3; $0.15/$0.50 per 1M tokens at list price
Technical Specifications
Key Features
Capabilities
Other Z.ai Models
Explore more models from Z.ai
GLM-5.3
Z.ai's flagship on GLM-5.2's base, with every gain from scaled-up post-training — frontier-level coding and cyber-defense capability
GLM-5.2
Zhipu AI's MIT-licensed flagship extending GLM-5.1 to a 1M-token context, the highest-ranked open-source model on long-horizon coding benchmarks
GLM-5.1
Zhipu AI's current flagship refined from GLM-5 for agentic engineering, topping the SWE-Bench Pro leaderboard with sustained 8-hour autonomous execution
