GLM-5.3-Flash

Zhipu AI

Parameters

320.0B total / 18.0B active

MoE: total / active

Architecture

Hybrid sparse + linear attention MoE with mHC

Released

27.08.2026

License

MIT License

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

Input Modalities

text image

Output Modalities

text

Context (native)

1,000,000 tokens

Context (extended)

1,000,000 tokens

Openness Index Score 100.0/100

About

GLM-5.3-Flash (zai-org/GLM-5.3-Flash) is the first natively multimodal model in the GLM-5 series - 320B total parameters with just 18B active, released August 27, 2026 under the MIT License. It outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks.

GLM-5.3-Flash starts from a newly trained base model with architecture and training recipe redesigned around capability and efficiency. For the first time in the GLM series it introduces a hybrid architecture combining sparse and linear attention, sharply reducing long-context serving costs while preserving precise long-context capabilities, and adopts Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. Together with its latest 30T-token multimodal pre-training corpus, these changes let GLM-5.3-Flash deliver more intelligence with less compute. Context length 1,000,000 tokens; text and image input. Blog: z.ai/blog/glm-5.3-flash; GLM-5 technical report: arXiv 2602.15763.

Training Data 30T-token multimodal pre-training corpus

Benchmark Scores

Benchmark Score Date Standing
Terminal Bench 2.1
coding_agent
93.02%
27.08.2026
#12/44
better than 74%
Agents' Last Exam
general_agent
48.16%
27.08.2026
#13/29
better than 57%
HLE (with tools)
stem_reasoning
73.80%
27.08.2026
#16/49
better than 69%
GDPVal-AA v2
general_agent
96.83%
27.08.2026
#2/38
better than 97%
MMLU
knowledge
84.57%
27.08.2026
#12/17
better than 31%
BIG-Bench Hard
reasoning
100.00%
27.08.2026
#1/8
better than 100%
Artificial Analysis Intelligence Index
composite
100.00%
27.08.2026
#1/3
better than 100%
HellaSwag
reasoning
68.32%
27.08.2026
#3/7
better than 67%
DeepSWE 1.1
coding_agent
82.29%
27.08.2026
#13/34
better than 64%
Automation-Bench
general_agent
73.08%
27.08.2026
#6/27
better than 81%

What is a good Terminal Bench 2.1 score? What is a good Agents' Last Exam score? What is a good HLE (with tools) score?

Hardware Requirements

Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 18.0B are active per token

Quantization Weights KV Cache Total VRAM On your hardware
FP16 — 16-bit 640.0 GB 2.8 GB 674.8 GB Which GPUs can run it? Calculate on your GPU
INT8 — 8-bit 320.0 GB 1.4 GB 337.4 GB Which GPUs can run it? Calculate on your GPU
INT4 — 4-bit 160.0 GB 0.7 GB 168.7 GB Which GPUs can run it? Calculate on your GPU
FP4 — 4-bit 160.0 GB 0.7 GB 168.7 GB Which GPUs can run it? Calculate on your GPU

Only datacenter GPUs — FP16 needs 674.8 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.

Architecture

Decoder Block ×45 input Embedding vocab 155K · d 4096 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×3 Full Attention 64 heads ×1 Linear / Recurrent 64 heads ×1 MoE FFN 288 experts · top-8 · +1 shared · dᴻ 2048 MTP Head ×1 speculative layer Final Norm LM Head vocab 155K output
Attention
Sparse Attention
MoE
288 experts · top-8 per token
Layers
45
Hidden size
4096
Context
1M tokens
Parameters
320000M
Active params
18000M

Source: Hugging Face config.json · Glm5NextForConditionalGeneration · exact layer pattern · model repo

Type: Mixture-of-Experts (MoE) with hybrid sparse + linear attention
Attention: Hybrid sparse + linear attention
Decoder: Transformer
MoE: yes (? experts)
Routing: top-k
Architecture

Hybrid sparse + linear attention with Manifold-Constrained Hyper-Connections (mHC)

Comparison

GLM-4.5: 355B total / 32B active / 92 layers; GLM-5.3-Flash: 320B total / 18B active / 45 layers

Key Innovation

First GLM model with hybrid sparse + linear attention, reducing long-context serving costs

Training Pipeline

  1. 1
    pretraining

    Multimodal Pre-training

    30T-token multimodal pre-training corpus. First GLM model trained with hybrid sparse + linear attention architecture and Manifold-Constrained Hyper-Connections (mHC).

  2. 2
    other

    Anonymous testing as ox-alpha

    Before release, GLM-5.3-Flash was tested anonymously as 'ox-alpha' on OpenCode and OpenRouter to gather user feedback. It became the most popular model of the week, with all traffic served on Chinese AI chips.

Training & Evaluation Datasets

NameRoleSizeModalitiesCollection
30T-token multimodal pre-training corpus pretraining 30T tokens text, image mixed

Linked Resources

paper

GLM-5: from Vibe Coding to Agentic Engineering

https://arxiv.org/abs/2602.15763

 blog

GLM-5.3-Flash: Frontier Intelligence, Flash Cost

https://z.ai/blog/glm-5.3-flash

 website

Z.ai API Platform - GLM-5.3-Flash

https://docs.z.ai/guides/llm/glm-5.3-flash

 github

SGLang - GLM-5.3-Flash cookbook

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

 github

vLLM - GLM-5.3-Flash recipes

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

 github

TokenSpeed - GLM-5.3-Flash

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

Comments (0)

No comments yet

Be the first to share your thoughts!

What now?

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

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