GLM 5.2

Layer3Labs' guide separates Zhipu's self-reported GLM 5.2 benchmark figures from independently confirmed results, highlighting that GLM 5.2's #1 finish on Design Arena's Code Categories leaderboard is the strongest fully independent result, with the model ranking roughly 10 Elo points ahead of Claude Fable 5 in a blind The specific point scores on SWE-bench Pro (62.1), Terminal-Bench 2.1, and FrontierSWE come from Zhipu's technical report and model card and are directionally corroborated by independent evaluators but have not been reproduced exactly by a third party using Zhipu's precise methodology. The guide frames the headline cla

GLM-5.2 is an open-source large language model from Z.ai (formerly Zhipu AI), released June 13, 2026, as the fourth model in the GLM-5 generation following GLM-5 (February 2026), GLM-5-Turbo (March 2026), and GLM-5.1 (April 2026). The model uses a Mixture-of-Experts Transformer architecture with 744 billion total param On academic benchmarks, GLM-5.2 scores 91.72% on MMLU and 80.63% on MMLU-Pro, both ahead of GLM-5.1. On coding, it reaches 81.0% on Terminal-Bench 2.1, up from GLM-5.1's 62.0%, with no parameter increase. The context window reaches 1 million tokens, five times GLM-5.1's, using DeepSeek Sparse Attention in layers 4 thro

Semgrep's June 22, 2026 security research evaluated several open-source models against their internal IDOR benchmark—the same dataset and prompt used to assess frontier coding agents. GLM-5.2 was the standout among models given only a prompt, reportedly beating Claude Opus 4.8 in this head-to-head, a notable result for This finding reinforces GLM-5.2's agentic coding and vulnerability-discovery capabilities independently of general intelligence benchmarks like Artificial Analysis. The benchmark is proprietary and single-axis (IDOR-focused), so it shouldn't be overgeneralized, but it provides concrete third-party evidence of strong se

Z.ai released the full benchmark scorecard for GLM 5.2 on June 17, 2026, four days after the June 13 launch, and simultaneously dropped the MIT-licensed open weights on HuggingFace for both zai-org/GLM-5.2 and the zai-org/GLM-5.2-FP8 quantized build. This was earlier than the originally announced "following week" guida By June 17, 24 community quantization variants were already published for llama.cpp, LM Studio, Jan, and Ollama, making GLM 5.2 the first credibly open-weight model to lead an Anthropic and OpenAI flagship on a real-world SWE-bench Pro head-to-head. The article frames the timing against Anthropic's Claude Agent SDK cre

GLM-5.2 was shipped by Zhipu AI (Z.ai) on June 13, 2026, as a 744-billion-parameter open-weight model under the MIT license, with 1 million-token context and 384 Mixture-of-Experts experts routing roughly 40 billion parameters per token. The article calls it text-only at launch with weights on HuggingFace (zai-org/GLM- On benchmarks, the article reports GLM-5.2 tops AIME 2026 at 99.2, reaches 81.0% on Terminal-Bench 2.1, and posts SWE-bench Pro 62.1 alongside competitive academic scores (91.72% MMLU, 80.63% MMLU-Pro). It uses DeepSeek-style IndexShare sparse attention claiming a 2.9x FLOP reduction at 1M context with improved Multi-T

A Hacker News thread (Jun 17, 2026, 916 points, 444 comments) discusses GLM-5.2's Artificial Analysis Intelligence Index v4.1 leadership at 51 among open-weights models. Community testers report practical token usage: GLM-5.2 averages ~42-43k output tokens per task on Artificial Analysis, compared to GPT-5.5 xhigh at 1 Developers note that running GLM-5.2 at "High" effort (rather than "Max") cuts token usage by roughly 2-2.5x with minimal quality loss for most tasks, reserving "Max" for complex problems. One user reported GLM-5.2 xhigh spending over 15 minutes and ~45k tokens before producing the first file on a 400-600 line Nim math

Z.ai released GLM-5.2 to coding plan subscribers on June 13, 2026, then published full open weights under an MIT license on June 16. According to Simon Willison's technical breakdown, the model is a 753B-parameter Mixture-of-Experts network with 40B active parameters, sized at 1.51TB, with a 1 million token context win Artificial Analysis rates GLM-5.2 at 51 on Intelligence Index v4.1, making it the leading open-weights model and ahead of Kimi K2.6 (43), DeepSeek V4 Pro max (44), and MiniMax-M3 (44), though output token usage rose to 43k per task (up from GLM-5.1's 26k). It ranks 2nd on the Code Arena WebDev leaderboard behind Claude

LLM-Stats ranks GLM-5.2 #28 on its composite LLM Stats Score as of September 19, 2026, with capability-tier standings placing Programming in the lower half at 5 of 7 with a 21.8 score (95% CI 22.8-26.7) and noting a 57-point gap between self-reported and real-task performance on the one comparable category. The tracker Performance degrades as conversation depth increases: 15.8 at Turn 1 (3 of 114), 15.3 across Turns 2-10 (11 of 115), 13.5 across Turns 11-30 (61 of 100), and 14.1 beyond Turn 31 (18 of 80), a drop of 1.7 points since turn 1. Category-specific scores show Chat at +1.49σ (280 votes), SVG at +1.38σ (42), Websites at +0.81