MODEL RELEASE · AUGUST 2026
GLM-5.3
Frontier Coding with Emergent Cyber Capabilities
Same base model
100% post-training
Weights in 2 weeks
THE IDEA
Same base model.
Every gain is post-training.
IndexShare
long-context
SAO
RL on long-horizon
slime
async training
0% base-model change
CODING
Massive coding gains
GLM-5.2 → GLM-5.3, same base, RL-scaled
EFFICIENCY
More output, fewer tokens
Completion rate vs. output tokens per task (Z.ai Code Bench)
EMERGENT CAPABILITY
It learned to exploit
Vulnerability discovery entered the training mix — capability kept scaling
REAL-WORLD TRANSFER
Real vulnerabilities, found
Tested against production codebases with security teams in China
0
vulnerabilities
0
projects
0
medium–high severity
0
years — oldest flaw (1981)
0
years avg. before discovery
INFRASTRUCTURE
Built on slime
Open-source RL-scaling framework — Megatron + SGLang
0
×
RL training throughput
1e-7
logprob consistency (99.99%↓)
300K
context window
USAGE
How to run it
Thinking effort
default
max · recommended for codingthinking.type: disabled
no longer supported
migrate to
enabled + low firstGLM Coding Plan
off-peak outside 14:00–18:00 UTC+8 weekdays
{ "model": "glm-5.3", "thinking": { "type": "enabled" }, "reasoning_effort": "max" }COMING SOON
Weights public in 2 weeks
Try GLM-5.3 in ZCode · Claude Code · OpenCode
z.ai
docs.z.ai/devpack
zcode.z.ai