GLM-5.2

On August 14, 2026, Z.ai published a research post on GLM-5.3, explicitly stating that GLM-5.3 uses the same base model as GLM-5.2 and that every reported gain over GLM-5.2 comes from post-training rather than a new base. The post describes scaling the GLM-5.2 stack: IndexShare for efficient long-context processing, SA The same post frames GLM-5.3 as the most capable open-weights coding model Z.ai has released, claiming a 50% improvement over GLM-5.2 on the in-house Z.ai Code Bench and open-source state-of-the-art results on Terminal Bench 3.0 and Agents' Last Exam. It also reports state-of-the-art performance on CyberGym for vulnera

Z.ai has officially introduced GLM-5.2, its flagship model purpose-built for long-horizon tasks and a substantial leap over GLM-5.1, delivering for the first time a solid 1M-token context window that the company says remains reliable under real engineering pressure. The model is released under an MIT open-source licens Architecturally, GLM-5.2 introduces IndexShare, which reuses the same indexer across every four sparse-attention layers to cut per-token FLOPs by 2.9× at a 1M context length, alongside an improved MTP layer that increases speculative-decoding acceptance length by up to 20%. The team substantially expanded 1M-context tr

Morph's August 21, 2026 article (last updated August 28, 2026) documents GLM-5.3 as Z.ai's August 14, 2026 upgrade of GLM-5.2, built on the same 753B mixture-of-experts base with no new pretraining — all gains from scaled post-training. Terminal-Bench 3.0 jumped from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Cyb GLM-5.3 ships as a post-training-only upgrade aimed at long-horizon agentic coding and cyber defense, with 1M context, 128K max output, and always-on reasoning (low/high/max effort, default max) that cannot be disabled. The Morph-deployed checkpoint incoai/GLM-5.3-NVFP4 is a ModelOpt NVFP4 expert quant with FP8 KV cach

Simon Willison's June 17, 2026 review notes that Z.ai released GLM-5.2 to coding plan subscribers on June 13 and then released full open weights under an MIT license on June 16, describing the 753B-parameter, 1.51TB MoE model (40 active parameters) as a text-input-only model with a 1-million-token context window, up fr The review highlights GLM-5.2 ranking 2nd on the Code Arena WebDev leaderboard (behind Claude Fable 5) for front-end web development including agentic coding workflows, and notes OpenRouter availability from 9 different providers at $1.40/M input and $4.40/M output, compared to GPT-5.5 at $5/$30 and Claude Opus 4.5-4.8

VentureBeat reported on June 16, 2026 that Z.ai released GLM-5.2, a 753-billion-parameter open-weights LLM targeted at long-horizon autonomous coding and engineering. According to the article, GLM-5.2 became available immediately on Hugging Face, the Z.ai API, and more than 20 third-party coding environments, with ente The article corroborates Z.ai's technical claims: a 1-million-token context window, the new IndexShare attention optimization that reuses one indexer across every four sparse attention layers and reduces per-token FLOPs by 2.9× at maximum context, and an upgraded Multi-Token Prediction layer that boosts speculative-dec