GLM-5.3

How it is built

GLM-5.3 starts from the same base model as GLM-5.2. Z.ai says every gain comes from post-training.¹¹1Z.ai, GLM-5.3 model card on Hugging Face: "the same base model as GLM-5.2", every gain from post-training; reasoning_effort defaults to max; benchmark table and footnotes (Terminal-Bench 3.0: Claude Code 2.1.207, avg@3; CyberGym: Claude Code 2.1.207, max reasoning effort, no web tools, single-run Pass@1 over 1,507 tasks). https://huggingface.co/zai-org/GLM-5.3 That base is a mixture of experts of 744B parameters, with 40B active for each token.²²2Z.ai, GLM-5 repository on GitHub, README, download table: GLM-5.3 at 744B-A40B in FP8, GLM-5.3-BF16 at 744B-A40B in BF16. https://github.com/zai-org/GLM-5. Hugging Face counts 753.33B parameters in the weight files; Z.ai does not explain the gap.

Of the 78 layers, the first 3 are dense. The other 75 carry 256 routed experts and 1 shared expert each, and the router keeps 8 per token.³³3GLM-5.3 config.json: num_hidden_layers 78, first_k_dense_replace 3, n_routed_experts 256, n_shared_experts 1, num_experts_per_tok 8, scoring_func sigmoid, topk_method noaux_tc, index_topk 2048, indexer_types with 21 "full" and 57 "shared" entries. https://huggingface.co/zai-org/GLM-5.3/blob/main/config.json. Arithmetic: 78 minus 3 = 75 MoE layers. It scores experts with a sigmoid. A bias on each score balances the load, with no auxiliary loss.³³

Attention is DeepSeek Sparse Attention: an indexer picks 2,048 earlier tokens for each query, and attention reads only those.³³ GLM-5.3 keeps IndexShare from GLM-5.2, which reuses one indexer across every four sparse attention layers.⁴⁴4Z.ai, GLM-5.2 model card on Hugging Face (IndexShare reuses the same indexer across every four sparse attention layers; paper arXiv 2603.12201). https://huggingface.co/zai-org/GLM-5.2

What is new

Thinking can no longer be switched off through the API. Effort runs from low to max, and max is the default.⁵⁵5Z.ai, GLM-5.3 API guide: 1M context, 128K maximum output, text-only input, reasoning always on with low, high and max effort. https://docs.z.ai/guides/llm/glm-5.3 ¹¹

On Terminal-Bench 3.0, averaged over three runs per task, GLM-5.2 scored 4.6. GLM-5.3 scores 28.3.¹¹ On CyberGym Z.ai reports 84.5, against 77.2 for GLM-5.2, in the Claude Code 2.1.207 harness at maximum effort with no web tools. Each of the 1,507 tasks got one attempt.¹¹

On 17 September 2026 the Center for AI Standards and Innovation (CAISI) at NIST called GLM-5.3 "the most cyber-capable open-weight model released to date".⁶⁶6NIST, "CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities", 17 September 2026. https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities Its evaluators also put the model about four months behind US frontier models on aggregate cyber tasks.⁶⁶

Running it

Z.ai announced GLM-5.3 on 14 August 2026 and opened it that day to its GLM Coding Plan and ZCode. API access and open weights were to follow in stages, after safety evaluations.⁷⁷7Z.ai on X, 14 August 2026 at 05:17 UTC: "GLM-5.3 is available now through GLM Coding Plan and ZCode. API access and open weights will be released in stages following rigorous safety evaluations." https://x.com/Zai_org/status/2088132969630212372. CAISI also gives 14 August as the release date. https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities The API opened on 18 August, as glm-5.3.⁸⁸8Z.ai, API release notes, GLM-5.3 entry dated 18 August 2026. https://docs.z.ai/release-notes/new-released It takes text only, with 1M tokens of context and up to 128K of output.⁵⁵ Input costs $1.40 per million tokens and output $4.40.⁹⁹9Z.ai, API pricing, USD per million tokens, read 29 September 2026: GLM-5.3 1.40 input, 0.26 cached input, 4.40 output. https://docs.z.ai/guides/overview/pricing

The FP8 weights on Hugging Face come to 755.6 GB, and the BF16 copy to 1,506.7 GB.²²¹⁰¹⁰10Sum of the .safetensors file sizes listed by Hugging Face, read 29 September 2026: zai-org/GLM-5.3, 141 files, 755.6 GB; zai-org/GLM-5.3-BF16, 282 files, 1,506.7 GB. https://huggingface.co/zai-org/GLM-5.3/tree/main and https://huggingface.co/zai-org/GLM-5.3-BF16/tree/main That is memory for the weights alone, before any KV cache.

The GLM-5.3 License grants MIT-style rights plus one clause for the largest users.¹¹¹¹11Z.ai, GLM-5.3 License, clause 2. https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE It covers a licensee whose group runs a Model as a Service business, such as inference sold through an API, and earns more than US$10 billion over any 12 consecutive months.¹¹¹¹ Such a licensee must pass Z.ai's security review before any commercial use.¹¹¹¹ GLM-5.3-Flash and GLM-5.2 stay under plain MIT.¹²¹²12MIT License files of GLM-5.3-Flash and GLM-5.2. https://huggingface.co/zai-org/GLM-5.3-Flash/blob/main/LICENSE and https://huggingface.co/zai-org/GLM-5.2