GLM-5.2

How it is built

GLM-5.2 keeps the size set by GLM-5: 744B parameters, with 40B active per token.¹¹1Z.ai, GLM-5.2 model card on Hugging Face: 1M-token context; IndexShare, per-token FLOPs cut by 2.9x at a 1M context (arXiv 2603.12201); MTP acceptance length up to 20% longer; multiple thinking effort levels; SWE-bench Pro 62.1 against 58.4 for GLM-5.1, run with OpenHands, temperature 1, top_p 1, max_new_tokens 32k, 400K context; MIT licence with "no regional limits". https://huggingface.co/zai-org/GLM-5.2 ²²2Z.ai, GLM-5 repository on GitHub, README: GLM-5 scales from 355B parameters (32B active) to 744B (40B active), pre-training data from 23T to 28.5T tokens, integrates DeepSeek Sparse Attention, "largely reducing deployment cost"; download table with GLM-5.2 in BF16 and GLM-5.2-FP8 at 744B-A40B. https://github.com/zai-org/GLM-5 GLM-5 had grown from the 355B of GLM-4.5, which ran 32B per token. It also raised pre-training data from 23T to 28.5T tokens.²²

Three dense layers open the stack of 78. The 75 below them are MoE layers with 256 routed experts and 1 shared expert, and a sigmoid router keeps 8 experts per token.³³3GLM-5.2 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, index_topk 2048, indexer_types with 21 "full" and 57 "shared" entries. https://huggingface.co/zai-org/GLM-5.2/blob/main/config.json. Arithmetic: 78 minus 3 = 75 MoE layers.

GLM-5 also brought in DeepSeek Sparse Attention, which Z.ai credits with a large cut in deployment cost.²² For each query an indexer chooses 2,048 earlier tokens, and attention reads those alone.³³

What is new

The context reached 1M tokens.⁴⁴4Z.ai, GLM-5.2 API guide: 1M context, 128K maximum output, text input and output. https://docs.z.ai/guides/llm/glm-5.2 GLM-5 and GLM-5.1 stopped at 202,752 positions.⁵⁵5config.json of GLM-5 and GLM-5.1, max_position_embeddings 202752. https://huggingface.co/zai-org/GLM-5/blob/main/config.json and https://huggingface.co/zai-org/GLM-5.1/blob/main/config.json

IndexShare lets one indexer serve four sparse attention layers.¹¹ In the config, 21 layers compute an indexer and 57 reuse one.³³ Z.ai puts the saving at 2.9 times fewer operations per token at a 1M context.¹¹

Z.ai also reworked the multi-token prediction layer, which drafts tokens for speculative decoding. The acceptance length grew by up to 20%.¹¹

On SWE-bench Pro Z.ai reports 62.1, against 58.4 for GLM-5.1. The run used OpenHands with a 400K context and up to 32K new tokens per reply.¹¹ Several thinking effort levels trade performance against latency.¹¹

GLM-5.3 was post-trained from this same base model, and the GLM-5.2 card now points to it as the newer version.⁶⁶6Z.ai, GLM-5.3 model card ("the same base model as GLM-5.2"), and the GLM-5.2 card metadata, new_version: zai-org/GLM-5.3-BF16. https://huggingface.co/zai-org/GLM-5.3 and https://huggingface.co/zai-org/GLM-5.2

Running it

glm-5.2 reached Z.ai's API on 16 June 2026.⁷⁷7Z.ai, API release notes, GLM-5.2 entry dated 16 June 2026. https://docs.z.ai/release-notes/new-released It reads text only. The context holds 1M tokens, and a reply up to 128K.⁴⁴ Prices match GLM-5.3: $1.40 per million input tokens and $4.40 per million output tokens.⁸⁸8Z.ai, API pricing, USD per million tokens, read 29 September 2026: GLM-5.2 and GLM-5.3 both 1.40 input, 0.26 cached input, 4.40 output. https://docs.z.ai/guides/overview/pricing

Z.ai publishes 1,506.7 GB of weights in BF16, and a second repository holds an FP8 copy of 755.6 GB.²²⁹⁹9Sum of the .safetensors file sizes listed by Hugging Face, read 29 September 2026: zai-org/GLM-5.2, 282 files, 1,506.7 GB; zai-org/GLM-5.2-FP8, 141 files, 755.6 GB. https://huggingface.co/zai-org/GLM-5.2/tree/main and https://huggingface.co/zai-org/GLM-5.2-FP8/tree/main

The licence is MIT, and the card adds that it has "no regional limits".¹¹ GLM-5.3 later moved to a licence of its own. It asks for a security review from Model as a Service groups with revenue above US$10 billion over 12 consecutive months.¹⁰¹⁰10Z.ai, GLM-5.3 License, clause 2. https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE