GLM-5.3

Zhipu AI

Parameters

753.3B total / 43.0B active

MoE: total / active

Architecture

MoE with IndexShare sparse attention (DSA) + MTP layer

Released

25.08.2026

License

GLM-5.3 License

Open Weights Commercial Use Multimodal BF16, F8_E4M3, F32 GLM en zh

Input Modalities

text

Output Modalities

text

Context (native)

1,048,576 tokens

Context (extended)

1,048,576 tokens

Openness Index Score 70.0/100

About

GLM-5.3 (zai-org/GLM-5.3) is Z.ai's flagship open-weights coding and cyber-capable model - the same 753B-total / 43B-activated MoE base model as GLM-5.2 (IndexShare sparse attention (DSA) + Multi-Token Prediction layer, 1,048,576-token context), where every gain comes from post-training targeting complex coding and long-horizon agentic tasks. Released August 25, 2026.

It is the most capable open-weights model for coding: a 50% improvement over GLM-5.2 on the in-house Z.ai Code Bench, and open-source SOTA on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam. As post-training scaled, an emergent cyber capability developed faster than expected: GLM-5.3 is state of the art on CyberGym for vulnerability discovery, and its gains are largest further up the exploitation chain, more than doubling GLM-5.2 on exploitation benchmarks (ExploitGym, ExploitBench). Thinking budget is controllable via the reasoning_effort parameter (low/high/max, default max), and the chat template's clear_thinking defaults to false for chat scenarios to pass explicitly.

Training Data Same base model as GLM-5.2; every gain comes from post-training targeting complex coding and long-horizon agentic tasks

Benchmark Scores

Benchmark Score Date Standing
Business Operation
agentic
33.56%
07.11.2026
#2/3
better than 50%
Terminal-Bench 4.0
coding_agent
47.65%
19.10.2026
#7/10
better than 33%
HANDBOOK.md
general_agent
20.80
07.11.2026
#3/3
better than 0%
HLE-Diamond (no tools)
stem_reasoning
19.66%
07.11.2026
#3/4
better than 33%
Agents' Last Exam (CLI)
general_agent
29.40
07.11.2026
#3/3
better than 0%
Frontier Finance
domain_finance
73.17%
07.11.2026
#2/3
better than 50%
Kimi Code Bench 3.0 Long-horizon
coding_agent
45.50
07.11.2026
#3/3
better than 0%
GDPval-AA v2.1
general_agent
1650.00
07.11.2026
#3/3
better than 0%
Automation Bench-AA
agentic
62.20
07.11.2026
#3/3
better than 0%
AA-Briefcase v1.1
general_agent
1507.00
07.11.2026
#3/3
better than 0%

Showing 10 of 74 results

What is a good Business Operation score? What is a good Terminal-Bench 4.0 score? What is a good HANDBOOK.md score?

Hardware Requirements

Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 43.0B are active per token

Quantization Weights KV Cache Total VRAM On your hardware
FP16 — 16-bit 1506.7 GB 6.7 GB 1588.7 GB Which GPUs can run it? Calculate on your GPU
INT8 — 8-bit 753.3 GB 3.3 GB 794.3 GB Which GPUs can run it? Calculate on your GPU
INT4 — 4-bit 376.7 GB 1.7 GB 397.2 GB Which GPUs can run it? Calculate on your GPU
FP4 — 4-bit 376.7 GB 1.7 GB 397.2 GB Which GPUs can run it? Calculate on your GPU

Only datacenter GPUs — FP16 needs 1588.7 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.

Architecture

Decoder Block input Embedding vocab 155K · d 6144 Full Attention 64 heads · dₕ 192 ×78 MoE FFN 256 experts · top-8 · +1 shared · dᴻ 2048 MTP Head ×1 speculative layer Final Norm LM Head vocab 155K output
Attention
Sparse Attention
MoE
256 experts · top-8 per token
Layers
78
Hidden size
6144
Context
1M tokens
Parameters
753329.9M
Active params
43046.5M

Source: Hugging Face config.json · GlmMoeDsaForCausalLM · model repo

Type: Sparse MoE Transformer with DeepSeek-style Sparse Attention (DSA) lightning indexer
Attention: MLA (q_lora_rank=2048, kv_lora_rank=512, qk_nope 192, qk_rope 64, v 256, 64 heads) + IndexShare lightning indexer (32 heads x 128-dim, index_topk=2048, re-index every 4 layers)
Decoder: Transformer decoder, 78 layers (3 dense FFN + 75 sparse MoE), SwiGLU
MoE: yes (256 experts)
Routing: sigmoid scoring (noaux_tc), top-8 + 1 shared expert, routed_scaling_factor 2.5

Training Pipeline

  1. 2
    rl

    Agentic post-training (coding + long-horizon/cyber RL)

    GLM-5.3 reuses the GLM-5.2 base model; all capability gains come from scaled post-training. Key outcomes: strongest open-weights coding model (+50% over GLM-5.2 on in-house Z.ai Code Bench; open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam), emergent cyber capability (SOTA on CyberGym vulnerability discovery; more than doubles GLM-5.2 on exploitation benchmarks ExploitGym/ExploitBench). Supports reasoning_effort low/high/max (default max) and clear_thinking.

Training & Evaluation Datasets

NameRoleSizeModalitiesCollection
Post-training corpus (coding, long-horizon agentic, cyber) rl — —

Linked Resources

arxiv

GLM-5: from Vibe Coding to Agentic Engineering (technical report, arXiv:2602.15763)

https://arxiv.org/abs/2602.15763

 github

zai-org/GLM-5 — GLM-5 & GLM-5.3 code, examples and resources

https://github.com/zai-org/GLM-5

 website

Transformers docs: GlmMoeDsa (GLM-5.3 architecture)

https://github.com/huggingface/transformers/blob/main/docs/source/en/model_doc/glm_moe_dsa.md

 website

SGLang cookbook: GLM-5.3

https://cookbook.sglang.io/autoregressive/GLM/GLM-5.3

 website

vLLM recipes: GLM-5.3

https://recipes.vllm.ai/zai-org/GLM-5.3

 website

TokenSpeed recipe: GLM-5.3

https://lightseek.org/tokenspeed/recipes/models#glm-5-3

Comments (0)

No comments yet

Be the first to share your thoughts!

What now?

You have read the full card for GLM-5.3 — here is where to go next.

Compare with another model Test on your own GPU Related models Data as JSON