GLM-5.3 Released as Open-Weight Model, Claims SOTA Coding Performance
Summary
- • GLM-5.3 released open-weight; all gains from post-training on the same GLM-5.2 base model
- • Claims open-weight SOTA in coding: 50% gain over GLM-5.2 on Z.ai Code Bench
- • Scores 28.3 on Terminal Bench 3.0, beating Kimi K3 (17.4) and Opus 4.8 (21.1)
- • Emergent cyber capability exceeded expectations; SOTA on CyberGym exploitation benchmarks
Details
GLM-5.3 open-weight release by zai-org
Released on HuggingFace as zai-org/GLM-5.3; same base model as GLM-5.2, all improvements from post-training. Available via Transformers, vLLM, SGLang, and Docker Model Runner.
50% coding gain over GLM-5.2 on Z.ai Code Bench
Scores 50% higher than GLM-5.2 on the internal Z.ai Code Bench and achieves open-weight SOTA on Terminal Bench 3.0 and Agents' Last Exam.
Terminal Bench 3.0: 28.3 vs. competitors
GLM-5.3 scores 28.3 vs. GLM-5.2's 4.6, Kimi K3's 17.4, and Opus 4.8's 21.1. GPT-5.6 Sol leads at 34.6, Fable 5 at 33.7 (with fallback).
DeepSWE v1.1: 66.9 vs. GLM-5.2's 46.2
Scores 66.9 on DeepSWE v1.1, ahead of Qwen3.8-Max (56.6), Opus 4.8 (58.0), and DeepSeek-V4 Pro (62.7); Kimi K3 leads at 67.5.
Unexpected cyber exploitation capability emergence
Cyber capability developed faster than expected during post-training. GLM-5.3 is state-of-the-art on CyberGym vulnerability discovery and more than doubles GLM-5.2's score on exploitation benchmarks.
Architecture: glm_moe_dsa with fp8 support
Uses the glm_moe_dsa mixture-of-experts architecture with fp8 quantization support; 4-item collection with 26 model variants on HuggingFace.
GLM-5.3 benchmark comparisons and capability highlights from the zai-org HuggingFace model card.
What This Means
GLM-5.3's open-weight release makes a top-tier coding model freely deployable, adding strong competition in a field that has grown increasingly crowded. Its benchmark results on Terminal Bench 3.0 and DeepSWE position it among the best open-weight coding models available today. The unexpected and faster-than-anticipated emergence of cyber exploitation capabilities during post-training is a notable safety signal—illustrating that capability jumps can occur in ways developers don't fully predict, raising questions about the adequacy of pre-release evaluation. For developers, GLM-5.3 offers a powerful freely accessible option that rivals commercial frontier models on a range of coding tasks.
Sources
- GLM-5.3 is now open-weightHugging Face