GLM-5.1

Z.ai's GLM-5.2 announcement dated June 16, 2026 explicitly references GLM-5.1 as its predecessor and characterizes the upgrade as a substantial leap in long-horizon task capability, including the first solid 1M-token context. The post lists specific GLM-5.1-relative improvements in GLM-5.2, such as the IndexShare spars For GLM-5.1 specifically, the page establishes that GLM-5.1 lacked a reliable 1M context and that GLM-5.2's expanded 1M-context training targets coding-agent scenarios including large-scale implementation, automated research, performance optimization, and complex debugging. This makes the GLM-5.2 blog a credible, model

MorphLLM provides a detailed third-party technical overview of GLM-5.1, confirming the core architecture: a 754B-total Mixture-of-Experts model with 40B active parameters per token, 200K-token context window, 128K-131K max output, and DSA sparse attention. Released April 7, 2026 under the MIT license on Hugging Face (z The model runs on transformers, vLLM, SGLang, KTransformers, and xLLM serving stacks, and integrates with Claude Code or Cline agent frameworks. GLM-5.1 scored 58.4 on SWE-Bench Pro (SOTA at launch), 63.5 on Terminal-Bench 2.0, and shows roughly 28% coding improvement over GLM-5 from refined post-training and asynchron

InferenceX provides a technical overview distinguishing GLM-5 from its follow-up GLM-5.1. GLM-5 scales from 355B parameters (32B active) in GLM-4.5 to 744B parameters (40B active) with 28.5T pre-training tokens, released February 11, 2026 under MIT license. GLM-5.1 is described as the follow-up point release achieving Both models are served through the Z.ai API with a 200K context window and 128K maximum output. The technical deep-dive details that GLM-5 integrates DeepSeek Sparse Attention (DSA) to reduce deployment cost while preserving long-context capacity, alongside an asynchronous RL infrastructure called "slime" that decouple

NYU Shanghai's RITS library published a third-party recap describing GLM-5.1 as a 754-billion-parameter open-weight Mixture-of-Experts model released by Z.ai on April 7, 2026, licensed under MIT and designed for agentic engineering. The recap reports that GLM-5.1 reached the #1 position on SWE-Bench Pro at 58.4%, ahead The write-up frames GLM-5.1 as a Dynamic Sparse Attention MoE with roughly 40 billion active parameters per token, capable of autonomously sustaining coding tasks for up to eight hours across hundreds of iterations. It cites demonstrations including a complete Linux desktop system built over an eight-hour, 655-iteratio

Z.ai's official developer documentation release notes confirm GLM-5.1 was released on April 7, 2026 as a model designed for long-horizon tasks capable of working independently for up to 8 hours in a single run. The release notes describe it as enabling a full loop from planning and execution to iterative refinement and According to the official docs, GLM-5.1 achieves comprehensive capability alignment with Claude Opus 4.6 and was built with multi-turn SFT, RL, and a process-based training approach. The documentation contextualizes GLM-5.1 within the broader model lineup, showing subsequent releases including GLM-5.2 (June 16, 2026, w