GLM-5.3-Flash

How it is built

GLM-5.3-Flash comes from a base model trained anew, on a multimodal corpus of 30T tokens.¹¹1Z.ai, GLM-5.3-Flash model card on Hugging Face: 320B total, 18B active, newly trained base, hybrid sparse and linear attention "for the first time in the GLM series", mHC, 30T-token multimodal corpus; outperforms GLM-5.2 "at one-tenth the price". https://huggingface.co/zai-org/GLM-5.3-Flash It has 320B parameters, and 18B of them work on each token.¹¹

Flash has 45 layers, the first 3 of them dense. In the other 42 a token goes to 8 of 288 routed experts, next to 1 shared expert.²²2GLM-5.3-Flash config.json, text_config: num_hidden_layers 45, first_k_dense_replace 3, n_routed_experts 288, n_shared_experts 1, num_experts_per_tok 8; layer_types with 34 linear_attention and 11 deepseek_sparse_attention entries, full_attn_layers 3, 7, 11 up to 43 (counted from zero); vision_config depth 24. https://huggingface.co/zai-org/GLM-5.3-Flash/blob/main/config.json. Arithmetic: 45 minus 3 = 42 MoE layers; positions 3 to 43 from zero are the 4th to the 44th layer.

Attention is hybrid, a first for the GLM series.¹¹ Linear attention runs in 34 layers. DeepSeek Sparse Attention runs in the other 11, at every fourth position from the 4th layer to the 44th.²²

Z.ai adopted Manifold-Constrained Hyper-Connections on the residual path to improve scaling efficiency.¹¹ A vision encoder of 24 layers turns images and video into tokens for the language model.²²

What is new

Against GLM-5.3, Z.ai reports attention compute down by a factor of 3.0 and the KV cache by a factor of 4.4.³³3Z.ai, GLM-5.3-Flash API guide: 3.0x less attention compute and 4.4x smaller KV cache than GLM-5.3; 320B against 355B, 18B against 32B active, 45 against 92 layers for GLM-4.5; DeepSWE v1.1 63.4 against 46.2 and AutomationBench 48.8 against 26.2 for GLM-5.2; served on a large-scale cluster of Chinese AI chips; FlashX at 200 tokens/s; 1M context, 128K output; video, image, text and file input. https://docs.z.ai/guides/llm/glm-5.3-flash

Z.ai also sets it beside GLM-4.5. The totals are close, 320B against 355B, but Flash runs 18B active parameters where GLM-4.5 ran 32B, over 45 layers instead of 92.³³

The card says Flash outperforms GLM-5.2 across benchmarks at a tenth of the price.¹¹ On DeepSWE v1.1 it scores 63.4 against 46.2.³³ The two runs differ, though. Flash had 6 hours per task at temperature 0.95, and the GLM-5.2 figure comes from a 2-hour run at temperature 1.0.⁴⁴4GLM-5.3-Flash model card, footnotes: DeepSWE with mini-swe-agent, temperature 0.95, top_p 1.0, timeout 6h, 400K context; AutomationBench v1.0.6. The GLM-5.2 DeepSWE figure matches the one on the GLM-5.2 card, whose run used a 2h timeout and temperature 1.0. https://huggingface.co/zai-org/GLM-5.3-Flash and https://huggingface.co/zai-org/GLM-5.2 On AutomationBench v1.0.6 the scores are 48.8 and 26.2.³³⁴⁴

Z.ai says it served the model on a large cluster of Chinese AI chips, at a cost per token it calls comparable to mainstream NVIDIA GPUs.³³

Running it

glm-5.3-flash has been on Z.ai's API since 26 August 2026.⁵⁵5Z.ai, API release notes, GLM-5.3-Flash entry dated 26 August 2026. https://docs.z.ai/release-notes/new-released It costs $0.15 per million input tokens and $0.50 per million output tokens, where GLM-5.2 costs $1.40 and $4.40.⁶⁶6Z.ai, API pricing, USD per million tokens, read 29 September 2026: GLM-5.3-Flash 0.15 input, 0.03 cached input, 0.50 output; GLM-5.3-FlashX 0.37, 0.075, 1.25; GLM-5.2 1.40, 0.26, 4.40. https://docs.z.ai/guides/overview/pricing. Arithmetic: 0.15 / 1.40 = 0.11; 0.50 / 4.40 = 0.11. The faster glm-5.3-flashx costs $0.37 and $1.25, and Z.ai quotes 200 tokens per second for it.⁶⁶³³

Besides text, the model reads images and video. Files are accepted too. It writes only text.³³ A reply can run to 128K tokens within a context of 1M.³³

The weights are MIT-licensed.⁷⁷7GLM-5.3-Flash LICENSE (MIT), and the download table in the GLM-5 repository README (FP8 and BF16). https://huggingface.co/zai-org/GLM-5.3-Flash/blob/main/LICENSE and https://github.com/zai-org/GLM-5 In FP8 they fill 328.3 GB, in BF16 642.7 GB, under half of GLM-5.3 in either format.⁸⁸8Sum of the .safetensors file sizes listed by Hugging Face, read 29 September 2026: zai-org/GLM-5.3-Flash, 62 files, 328.3 GB; zai-org/GLM-5.3-Flash-BF16, 120 files, 642.7 GB; GLM-5.3 for comparison, 755.6 GB in FP8 and 1,506.7 GB in BF16. https://huggingface.co/zai-org/GLM-5.3-Flash/tree/main and https://huggingface.co/zai-org/GLM-5.3-Flash-BF16/tree/main. Arithmetic: 328.3 / 755.6 = 0.43; 642.7 / 1,506.7 = 0.43.