GLM 5.2

Z.ai (Zhipu AI) announced GLM-5.2 as its latest flagship model engineered for long-horizon tasks, published on its official blog dated June 16, 2026. The model delivers a solid 1-million-token context window that the vendor says holds up under sustained coding-agent workloads, and introduces a new architecture componen On three long-horizon coding benchmarks reported by Z.ai, GLM-5.2 trails Claude Opus 4.8 by only 1% on FrontierSWE while edging out GPT-5.5 by 1% and Opus 4.7 by 11%; on PostTrainBench, where each agent is given an H100 GPU for model post-training, GLM-5.2 ranks second only to Opus 4.8, outperforming both Opus 4.7 and

A Baidu Cloud technical analysis dated September 7, 2026 examines GLM-5.2's architecture and performance, confirming the model maintains the GLM-5 series' 744 billion total parameters with 40 billion activated parameters per token. The article attributes the model's long-context efficiency to a sparse attention mechani The Baidu analysis reports benchmark results including a 32% improvement in long-sequence reasoning scores versus GLM-5.1 in KingBench 3 testing and 91% code completion accuracy in cross-file reference scenarios, positioning GLM-5.2 as outperforming multimodal models limited to 256K context windows in complex engineeri

Featherless published a Day Zero launch breakdown on 2026-06-18 confirming that GLM-5.2 ships as a Mixture-of-Experts model roughly 753B parameters in size with about 39B parameters activated per token, released by Z.ai under an MIT license and exposed through an OpenAI-compatible API. The piece attributes the version- The same post lists concrete benchmark deltas attributed to GLM-5.2 versus predecessor GLM-5.1: Terminal-Bench 2.1 rose from 63.5 to 81.0, SWE-bench Pro from 58.4 to 62.1, FrontierSWE from 30.5 to 74.4, and SWE-Marathon from 1.0 to 13.0, with reasoning gains on AIME 2026 (95.3 to 99.2) and GPQA-Diamond (86.2 to 91.2).

A third-party review on aiforanything.io covers Zhipu AI's GLM-5.2, reporting a release date of June 13, 2026, and listing confirmed specs: a Mixture-of-Experts architecture with 744 billion total parameters and 40 billion active parameters per token, a 1,000,000-token input context window, a maximum output of 131,072 The review positions GLM-5.2 as the strongest open model for coding and agentic work, citing pricing roughly 10× cheaper than comparable frontier access (with a GLM Coding Plan starting at roughly $10/month), but flags that Zhipu shipped it without published benchmarks in the reviewer's view, meaning performance claims

Z.ai's official developer release-notes index explicitly lists GLM-5.2 as a 2026-06-16 entry, distinct from later GLM-5.3 (2026-08-18) and GLM-5.3-Flash (2026-08-26) entries, helping fix the exact version under discussion. The GLM-5.2 entry states the model supports 1M lossless context, significantly improving long-hor By enumerating GLM-5.1, GLM-5.2, GLM-5.3, and GLM-5.3-Flash as separate line items with their own dates, the release-notes page functions as a version discriminator that rules out inadvertently splicing sibling-family evidence onto GLM-5.2. The page is first-party Z.ai material, so creator attribution to Z.ai (Zhipu) i