GLM-5.2

Flexible Thinking Effort Levels

GLM-5.2 introduces effort level control, enabling users to explicitly balance model capability against task execution speed and computational cost.

Effort Levels

The model supports multiple thinking effort levels (e.g., low, medium, max) that control how much reasoning the model performs before generating a response.

Performance vs. Cost Trade-off

As shown in the GLM-5.2 blog, the model delivers substantially stronger agentic coding performance than GLM-5.1 at comparable token budgets. Its capability is roughly positioned between Claude Opus 4.7 and Claude Opus 4.8 under similar token consumption.

The Max effort level allows users to allocate additional computation when higher performance is required for challenging tasks, further extending the model's coding capability.

Evaluation Usage

• ProgramBench: Uses reasoning_effort=max with 400K context
• FrontierSWE: Uses max effort level with 1M context
• PostTrainBench: Uses max effort level with 1M context
• SWE-Marathon: Uses max effort level with 1M context

MIT Open Source License

GLM-5.2 is released under the MIT License — one of the most permissive open-source licenses available.

Key Terms

• Commercial use: Allowed (the model is marked as commercial_use_allowed: True in the DB)
• Open source: Fully open source (is_open_source: True)
• No regional limits: The license explicitly states "no regional limits, technical access without borders"
• Modification and redistribution: Permitted under MIT terms

Significance

The "Pure Open" positioning is a key differentiator for GLM-5.2. Unlike many frontier models that restrict usage by region or impose commercial limitations, GLM-5.2 provides unrestricted access under MIT, making it one of the most openly available frontier-class models.

Model Tree and Ecosystem

Model Tree for zai-org/GLM-5.2

Community Resources

• HuggingFace Collection: GLM-5.2 Collection — 2 items, updated Jun 16
• Blog Post: GLM-5.2: Built for Long-Horizon Tasks — Published Jun 17
• GitHub: zai-org/GLM-5
• API Platform: Z.ai API
• Chat Demo: chat.z.ai

Community

• Discord: Join
• WeChat: QR Code

Notable Quantizations

• unsloth/GLM-5.2-GGUF — GGUF format
• QuantTrio/GLM-5.2-Int4-Int8Mix — INT4/INT8 mixed quantization
• zai-org/GLM-5.2-FP8 — FP8 quantization (also on Dell Enterprise Hub)

Statistics

• Downloads (last month): 2,576,335
• Likes: 5.04K
• Followers: 19K

BibTeX Citation

If you find GLM-5.2 useful in your research, please cite the technical report:

@misc{glm5team2026glm5vibecodingagentic,
      title={GLM-5: from Vibe Coding to Agentic Engineering},
      author={GLM-5-Team and Aohan Zeng and Xin Lv and Zhenyu Hou and Zhengxiao Du and Qinkai Zheng and Bin Chen and Da Yin and Chendi Ge and Chenghua Huang and Chengxing Xie and Chenzheng Zhu and Congfeng Yin and Cunxiang Wang and Gengzheng Pan and Hao Zeng and Haoke Zhang and Haoran Wang and Huilong Chen and Jiajie Zhang and Jian Jiao and Jiaqi Guo and Jingsen Wang and Jingzhao Du and Jinzhu Wu and Kedong Wang and Lei Li and Lin Fan and Lucen Zhong and Mingdao Liu and Mingming Zhao and Pengfan Du and Qian Dong and Rui Lu and Shuang-Li and Shulin Cao and Song Liu and Ting Jiang and Xiaodong Chen and Xiaohan Zhang and Xuancheng Huang and Xuezhen Dong and Yabo Xu and Yao Wei and Yifan An and Yilin Niu and Yitong Zhu and Yuanhao Wen and Yukuo Cen and Yushi Bai and Zhongpei Qiao and Zihan Wang and Zikang Wang and Zilin Zhu and Ziqiang Liu and Zixuan Li and Bojie Wang and Bosi Wen and Can Huang and Changpeng Cai and Chao Yu and Chen Li and Chengwei Hu and Chenhui Zhang and Dan Zhang and Daoyan Lin and Dayong Yang and Di Wang and Ding Ai and Erle Zhu and Fangzhou Yi and Feiyu Chen and Guohong Wen and Hailong Sun and Haisha Zhao and Haiyi Hu and Hanchen Zhang and Hanrui Liu and Hanyu Zhang and Hao Peng and Hao Tai and Haobo Zhang and He Liu and Hongwei Wang and Hongxi Yan and Hongyu Ge and Huan Liu and Huanpeng Chu and Jia'ni Zhao and Jiachen Wang and Jiajing Zhao and Jiamin Ren and Jiapeng Wang and Jiaxin Zhang and Jiayi Gui and Jiayue Zhao and Jijie Li and Jing An and Jing Li and Jingwei Yuan and Jinhua Du and Jinxin Liu and Junkai Zhi and Junwen Duan and Kaiyue Zhou and Kangjian Wei and Ke Wang and Keyun Luo and Laiqiang Zhang and Leigang Sha and Liang Xu and Lindong Wu and Lintao Ding and Lu Chen and Minghao Li and Nianyi Lin and Pan Ta and Qiang Zou and Rongjun Song and Ruiqi Yang and Shangqing Tu and Shangtong Yang and Shaoxiang Wu and Shengyan Zhang and Shijie Li and Shuang Li and Shuyi Fan and Wei Qin and Wei Tian and Weining Zhang and Wenbo Yu and Wenjie Liang and Xiang Kuang and Xiangmeng Cheng and Xiangyang Li and Xiaoquan Yan and Xiaowei Hu and Xiaoying Ling and Xing Fan and Xingye Xia and Xinyuan Zhang and Xinze Zhang and Xirui Pan and Xu Zou and Xunkai Zhang and Yadi Liu and Yandong Wu and Yanfu Li and Yidong Wang and Yifan Zhu and Yijun Tan and Yilin Zhou and Yiming Pan and Ying Zhang and Yinpei Su and Yipeng Geng and Yong Yan and Yonglin Tan and Yuean Bi and Yuhan Shen and Yuhao Yang and Yujiang Li and Yunan Liu and Yunqing Wang and Yuntao Li and Yurong Wu and Yutao Zhang and Yuxi Duan and Yuxuan Zhang and Zezhen Liu and Zhengtao Jiang and Zhenhe Yan and Zheyu Zhang and Zhixiang Wei and Zhuo Chen and Zhuoer Feng and Zijun Yao and Ziwei Chai and Ziyuan Wang and Zuzhou Zhang and Bin Xu and Minlie Huang and Hongning Wang and Juanzi Li and Yuxiao Dong and Jie Tang},
      year={2026},
      eprint={2602.15763},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.15763}
}

Related Papers

• GLM-5 Technical Report: arXiv:2602.15763 — "GLM-5: from Vibe Coding to Agentic Engineering"
• IndexShare Paper: arXiv:2603.12201 — "IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse"

Evaluation Methodology

Reasoning Benchmarks

• HLE & other reasoning tasks: Sampling parameters: temperature=1.0, top_p=0.95. Maximum generation length: 163,840 tokens. Text-only subset reported by default; results marked with * are from the full set. Judge model: GPT-5.5 (medium). For HLE-with-tools: max context length 300,000 tokens, no context management strategy.
• AIME, HMMT, IMOAnswerBench: System prompt: Your response should be in the following format: Explanation: {your explanation} Exact Answer: {your succinct, final answer} Confidence: {your confidence score between 0% and 100%}. Judge: GPT-5.5 (medium).

Coding Benchmarks

• SWE-Bench Pro: Run with OpenHands using a tailored instruction prompt. Settings: temperature=1, top_p=1, max_new_tokens=32k, 400K context window.
• NL2Repo: temperature=1.0, top_p=1.0, max_new_tokens=48k, 400K context. Rule-based and LLM-based judgement to prevent malicious behaviors.
• DeepSWE: Official pier evaluation framework, mini-swe-agent harness. temperature=1.0, top_p=1.0, timeout=2h, 400K context. Isolated container: 2 CPUs, 8 GB RAM, no internet.
• ProgramBench: 200 instances, Claude-Code 2.1.156. temperature=1.0, top_p=1.0, max_tokens=64000, max_turns=2000, sample_timeout=6h, reasoning_effort=max, 400K context. Sandbox: 4 CPUs, 8 GB RAM, internet disabled.
• Terminal-Bench 2.1 (Terminus 2): Terminus-2 framework, parser=json, timeout=4h, temperature=1.0, top_p=1.0, max_new_tokens=48k, max_episodes=500, 256K context. Resources: 4 CPUs, 8 GB RAM.
• Terminal-Bench 2.1 (Claude Code): Claude Code 2.1.167, temperature=1.0, top_p=0.95, max_new_tokens=131072 (overridden to 128k via proxy). No wall-clock time limits. Scores averaged over 5 runs.
• FrontierSWE: Conducted by Proximal. 1M context, max effort, 128K max output tokens. Dominance score as of 2026/06/16.
• PostTrainBench: Conducted by PostTrainBench. 1M context, max effort, 128K max output tokens.
• SWE-Marathon: Conducted by Abundant AI. 1M context, max effort, 128K max output tokens.

Agentic Benchmarks

• MCP-Atlas: Think mode, 500-task public subset, 10-minute timeout per task. Judge: Gemini-3.0-Pro.
• Tool-Decathlon: Official evaluation service, max_token=128K.

Supported Inference Frameworks

GLM-5.2 supports deployment with the following inference frameworks:

GPU Deployment

Ascend NPU Deployment

For deployment on the Ascend NPU platform, the following inference frameworks are supported:

• vLLM-Ascend
• xLLM
• SGLang

Reference: Ascend deployment guide

GLM-5.2 Full Benchmark Results

All scores are for GLM-5.2. Comparative models shown for context.

Reasoning Benchmarks

Coding Benchmarks

Agentic Benchmarks

Key Takeaways

• GLM-5.2 significantly outperforms GLM-5.1 across all benchmark categories
• On Terminal Bench 2.1, GLM-5.2 (81.0) is within a few points of Claude Opus 4.8 (85.0)
• On DeepSWE, GLM-5.2 (46.2) is a massive improvement over GLM-5.1 (18.0)
• On FrontierSWE (Dominance), GLM-5.2 (74.4) is competitive with Claude Opus 4.8 (75.1)
• Scores marked with * are from the full HLE set, not text-only subset

Solid 1M-Token Context

GLM-5.2 delivers, for the first time in the GLM series, a solid 1M-token context that stably sustains long-horizon work.

Context Window

• Native context length: 1,000,000 tokens (1M)
• Extended context length: 1,000,000 tokens (1M) — no extension beyond native
• Stability: The context is described as "solid" — it stably sustains long-horizon tasks without degradation, a substantial improvement over GLM-5.1.

Significance

This 1M context is a core differentiator for GLM-5.2, enabling it to handle long-horizon agentic coding tasks, large document processing, and extended multi-turn reasoning that were not feasible with GLM-5.1's context window.

MTP Layer for Speculative Decoding

GLM-5.2 improves the Multi-Token Prediction (MTP) layer used for speculative decoding, achieving two key objectives:

Improvements over GLM-5.1

1. Increased acceptance length: The MTP layer improvements increase the speculative decoding acceptance length by up to 20% compared to GLM-5.1.
2. Training-inference consistency: GLM-5.2 eliminates the training-inference discrepancy present in GLM-5.1's MTP layer. In the two-step MTP inference, the first step is consistent with training (all hidden states from the target model). The second step resolves the mismatch where some hidden states come from the MTP layer rather than the target model.

How It Works

In a two-step MTP layer:

• Step 1: All hidden states come from the target model — inference is consistent with training.
• Step 2: Some hidden states come from the target model and some from the MTP layer, which previously caused a training-inference discrepancy in GLM-5.1.

IndexShare Sparse Attention

GLM-5.2 introduces IndexShare, a novel architectural innovation for sparse attention that reuses the same indexer across every four sparse attention layers.

Key Properties

• Cross-layer index reuse: The same indexer is shared across groups of four consecutive sparse attention layers, eliminating redundant indexer computations.
• FLOP reduction: Reduces per-token FLOPs by 2.9× at a 1M context length compared to non-shared indexing.
• Architecture tag: glm_moe_dsa (GLM Mixture-of-Experts with Dynamic Sparse Attention)

Related Papers

• IndexCache: Accelerating Sparse Attention via Cross-Layer Index Reuse — Published March 2026

Model Configuration

• Total Parameters: 753B
• Architecture Type: MoE with IndexShare sparse attention + MTP layer
• Tensor Types: BF16, F32

GLM-5.2 Introduction and Key Capabilities

GLM-5.2 is the latest flagship model from Zhipu AI (zai-org), designed for long-horizon tasks. It marks a substantial leap in long-horizon task capability over its predecessor GLM-5.1 and, for the first time, delivers that capability on a solid 1M-token context.

Key Capabilities

1. Solid 1M Context — A solid 1M-token context that stably sustains long-horizon work.
2. Advanced Coding with Flexible Effort — Stronger coding capabilities with multiple thinking effort levels to balance performance and latency.
3. Improved Architecture — IndexShare reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length. The MTP layer for speculative decoding is improved, increasing acceptance length by up to 20%.
4. Pure Open — An MIT open-source license with no regional limits and technical access without borders.

Model Identity

• Provider: Zhipu AI (zai-org)
• Model Family: GLM
• Release Date: 2026-06-16
• Parameters: 753B
• Tensor Type: BF16, F32
• Languages: English, Chinese
• Tags: Text Generation, Transformers, Safetensors, glm_moe_dsa, conversational, Eval Results