The GLM-5.2 model card is the README of the zai-org/GLM-5.2 repository on Hugging Face, published by Z.ai on June 16, 2026 with the weights of GLM-5.2 (Source: huggingface.co). It is not a system card: it contains no safety evaluations, and for architecture and training it points to the GLM-5 technical report. Its evaluation content is one benchmark table and a set of footnotes describing how the figures were produced. The footnotes were added to the README on June 23, 2026.
Model description
The card describes GLM-5.2 as Z.ai's flagship model for long-horizon tasks and lists four changes from GLM-5.1: a 1M-token context; stronger coding with several thinking-effort levels; an architecture change the card calls IndexShare, which reuses one indexer across every four sparse-attention layers and which Z.ai says reduces per-token FLOPs by 2.9× at a 1M context length, together with a revised multi-token-prediction layer that Z.ai says raises speculative-decoding acceptance length by up to 20%; and release under the MIT licence without regional limits. It lists supported inference frameworks (SGLang, vLLM, Transformers, KTransformers, Unsloth, and Ascend NPU back ends).
Evaluation approach
The benchmark table has eight columns — GLM-5.2, GLM-5.1, Qwen3.7-Max, MiniMax M3, DeepSeek-V4-Pro, Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro — and three groups of rows.
- Reasoning: Humanity's Last Exam with and without tools, CritPt, AIME 2026, HMMT November 2025 and February 2026, IMOAnswerBench and GPQA-Diamond. Z.ai reports sampling at temperature 1.0 and top-p 0.95 with up to 163,840 generated tokens and GPT-5.5 (medium) as the judge model. HLE figures are for the text-only subset unless marked with an asterisk, in which case they come from the full set; the asterisked entries are those for Claude Opus 4.8 and GPT-5.5, and for Gemini 3.1 Pro with tools. HLE with tools was run with a 300,000-token context and no context management.
- Coding: SWE-bench Pro (OpenHands, 400K context), NL2Repo, DeepSWE (mini-swe-agent, two-hour timeout, no internet), ProgramBench (200 instances in Claude Code 2.1.156), Terminal-Bench 2.1 under the Terminus-2 framework and under the best reported harness, FrontierSWE, PostTrainBench and SWE-Marathon. The card states that the last three were run by outside parties — Proximal, PostTrainBench and Abundant AI respectively — at 1M context and maximum effort, and that the FrontierSWE figure is a Dominance score as of June 16, 2026.
- Agentic: MCP-Atlas on the 500-task public subset with a 10-minute timeout and Gemini 3.0 Pro as judge, and Tool-Decathlon through its official evaluation service.
The card does not say which of the comparison models' figures Z.ai measured itself and which it took from other developers' publications, except where a footnote states that all models were evaluated under one configuration (MCP-Atlas).
Reported results
Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.
| Evaluation | Model | Score | Setting | Section |
|---|---|---|---|---|
| CritPt | GLM-5.2 | 20.9% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; developer-reported | Benchmark; Footnote |
| CritPt | GLM-5.1 | 4.6% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; developer-reported | Benchmark; Footnote |
| CritPt | Qwen3.7-Max | 13.4% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| CritPt | MiniMax M3 | 3.7% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| CritPt | DeepSeek-V4-Pro | 12.9% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| CritPt | Claude Opus 4.8 | 20.9% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| CritPt | GPT-5.5 | 27.1% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| CritPt | Gemini 3.1 Pro | 17.7% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| IMOAnswerBench | GLM-5.2 | 91.0% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; fixed Explanation / Exact Answer / Confidence system prompt; GPT-5.5 (medium) as judge; developer-reported | Benchmark; Footnote |
| IMOAnswerBench | GLM-5.1 | 83.8% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; fixed Explanation / Exact Answer / Confidence system prompt; GPT-5.5 (medium) as judge; developer-reported | Benchmark; Footnote |
| IMOAnswerBench | Qwen3.7-Max | 90% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; fixed Explanation / Exact Answer / Confidence system prompt; GPT-5.5 (medium) as judge; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| IMOAnswerBench | DeepSeek-V4-Pro | 89.8% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; fixed Explanation / Exact Answer / Confidence system prompt; GPT-5.5 (medium) as judge; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| IMOAnswerBench | Claude Opus 4.8 | 83.5% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; fixed Explanation / Exact Answer / Confidence system prompt; GPT-5.5 (medium) as judge; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| IMOAnswerBench | Gemini 3.1 Pro | 81% | temperature 1.0, top_p 0.95, maximum generation length 163,840 tokens; fixed Explanation / Exact Answer / Confidence system prompt; GPT-5.5 (medium) as judge; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| NL2Repo | GLM-5.2 | 48.9% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; developer-reported | Benchmark; Footnote |
| NL2Repo | GLM-5.1 | 42.7% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; developer-reported | Benchmark; Footnote |
| NL2Repo | Qwen3.7-Max | 47.2% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| NL2Repo | MiniMax M3 | 42.1% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| NL2Repo | DeepSeek-V4-Pro | 35.5% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| NL2Repo | Claude Opus 4.8 | 69.7% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| NL2Repo | GPT-5.5 | 50.7% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| NL2Repo | Gemini 3.1 Pro | 33.4% | temperature 1.0, top_p 1.0, max_new_tokens 48k, 400K context; rule-based and LLM-based checks against unauthorized pip or curl operations; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| DeepSWE | GLM-5.2 | 46.2% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; developer-reported | Benchmark; Footnote |
| DeepSWE | GLM-5.1 | 18% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; developer-reported | Benchmark; Footnote |
| DeepSWE | Qwen3.7-Max | 18% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| DeepSWE | MiniMax M3 | 20% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| DeepSWE | DeepSeek-V4-Pro | 8% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| DeepSWE | Claude Opus 4.8 | 58% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| DeepSWE | GPT-5.5 | 70% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| DeepSWE | Gemini 3.1 Pro | 10% | official pier evaluation framework with the mini-swe-agent harness; temperature 1.0, top_p 1.0, 2h timeout, 400K context; isolated container with 2 CPUs, 8 GB RAM, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| ProgramBench | GLM-5.2 | 63.7% | 200 instances; Claude Code 2.1.156; temperature 1.0, top_p 1.0, max_tokens 64,000, max_turns 2,000, 6h sample timeout, reasoning_effort max, 400K context; 4 CPUs / 8 GB RAM sandbox, no internet; developer-reported | Benchmark; Footnote |
| ProgramBench | GLM-5.1 | 50.9% | 200 instances; Claude Code 2.1.156; temperature 1.0, top_p 1.0, max_tokens 64,000, max_turns 2,000, 6h sample timeout, reasoning_effort max, 400K context; 4 CPUs / 8 GB RAM sandbox, no internet; developer-reported | Benchmark; Footnote |
| ProgramBench | DeepSeek-V4-Pro | 47.8% | 200 instances; Claude Code 2.1.156; temperature 1.0, top_p 1.0, max_tokens 64,000, max_turns 2,000, 6h sample timeout, reasoning_effort max, 400K context; 4 CPUs / 8 GB RAM sandbox, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| ProgramBench | Claude Opus 4.8 | 71.9% | 200 instances; Claude Code 2.1.156; temperature 1.0, top_p 1.0, max_tokens 64,000, max_turns 2,000, 6h sample timeout, reasoning_effort max, 400K context; 4 CPUs / 8 GB RAM sandbox, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| ProgramBench | GPT-5.5 | 70.8% | 200 instances; Claude Code 2.1.156; temperature 1.0, top_p 1.0, max_tokens 64,000, max_turns 2,000, 6h sample timeout, reasoning_effort max, 400K context; 4 CPUs / 8 GB RAM sandbox, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| ProgramBench | Gemini 3.1 Pro | 39.5% | 200 instances; Claude Code 2.1.156; temperature 1.0, top_p 1.0, max_tokens 64,000, max_turns 2,000, 6h sample timeout, reasoning_effort max, 400K context; 4 CPUs / 8 GB RAM sandbox, no internet; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| FrontierSWE (Dominance) | GLM-5.2 | 74.4% | evaluation conducted by Proximal; 1M context, max effort, 128K maximum output tokens; Dominance score as of 2026-06-16; developer-reported | Benchmark; Footnote |
| FrontierSWE (Dominance) | GLM-5.1 | 30.5% | evaluation conducted by Proximal; 1M context, max effort, 128K maximum output tokens; Dominance score as of 2026-06-16; developer-reported | Benchmark; Footnote |
| FrontierSWE (Dominance) | DeepSeek-V4-Pro | 29.0% | evaluation conducted by Proximal; 1M context, max effort, 128K maximum output tokens; Dominance score as of 2026-06-16; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| FrontierSWE (Dominance) | Claude Opus 4.8 | 75.1% | evaluation conducted by Proximal; 1M context, max effort, 128K maximum output tokens; Dominance score as of 2026-06-16; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| FrontierSWE (Dominance) | GPT-5.5 | 72.6% | evaluation conducted by Proximal; 1M context, max effort, 128K maximum output tokens; Dominance score as of 2026-06-16; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| FrontierSWE (Dominance) | Gemini 3.1 Pro | 39.6% | evaluation conducted by Proximal; 1M context, max effort, 128K maximum output tokens; Dominance score as of 2026-06-16; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| PostTrainBench | GLM-5.2 | 34.3% | evaluation conducted by PostTrainBench; 1M context, max effort, 128K maximum output tokens; developer-reported | Benchmark; Footnote |
| PostTrainBench | GLM-5.1 | 20.1% | evaluation conducted by PostTrainBench; 1M context, max effort, 128K maximum output tokens; developer-reported | Benchmark; Footnote |
| PostTrainBench | Claude Opus 4.8 | 37.2% | evaluation conducted by PostTrainBench; 1M context, max effort, 128K maximum output tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| PostTrainBench | GPT-5.5 | 28.4% | evaluation conducted by PostTrainBench; 1M context, max effort, 128K maximum output tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| PostTrainBench | Gemini 3.1 Pro | 21.6% | evaluation conducted by PostTrainBench; 1M context, max effort, 128K maximum output tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| SWE-Marathon | GLM-5.2 | 13.0% | evaluation conducted by Abundant AI; 1M context, max effort, 128K maximum output tokens; developer-reported | Benchmark; Footnote |
| SWE-Marathon | GLM-5.1 | 1.0% | evaluation conducted by Abundant AI; 1M context, max effort, 128K maximum output tokens; developer-reported | Benchmark; Footnote |
| SWE-Marathon | Claude Opus 4.8 | 26.0% | evaluation conducted by Abundant AI; 1M context, max effort, 128K maximum output tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| SWE-Marathon | GPT-5.5 | 12.0% | evaluation conducted by Abundant AI; 1M context, max effort, 128K maximum output tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| SWE-Marathon | Gemini 3.1 Pro | 4.0% | evaluation conducted by Abundant AI; 1M context, max effort, 128K maximum output tokens; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| Tool-Decathlon | GLM-5.2 | 48.2% | official evaluation service; max_token 128K; developer-reported | Benchmark; Footnote |
| Tool-Decathlon | GLM-5.1 | 40.7% | official evaluation service; max_token 128K; developer-reported | Benchmark; Footnote |
| Tool-Decathlon | DeepSeek-V4-Pro | 52.8% | official evaluation service; max_token 128K; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| Tool-Decathlon | Claude Opus 4.8 | 59.9% | official evaluation service; max_token 128K; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| Tool-Decathlon | GPT-5.5 | 55.6% | official evaluation service; max_token 128K; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
| Tool-Decathlon | Gemini 3.1 Pro | 48.8% | official evaluation service; max_token 128K; reported by Z.ai (Zhipu AI) in the GLM-5.2 model card | Benchmark; Footnote |
Provenance
The raw file is the verbatim README fetched from Hugging Face on September 19, 2026; its last edit, on September 1, 2026, added a pointer to GLM-5.3 as the newer version. A benchmark chart embedded as an image is not reproduced. All results are developer-reported or relayed by the developer, which is the basis for the medium confidence rating. The same table appears in Z.ai's launch post (Source: z.ai).
Relationships
- supports: GLM-5.2 — primary reference for the model's reported benchmark results.
- depends-on: GLM-5: from Vibe Coding to Agentic Engineering (Zhipu AI, 2026) — the technical report the card cites for the GLM-5 series.
- related: GLM-5.3 Model Card (Z.ai, August 2026), Zhipu AI