GLM-5 and GLM-5.1

GLM-5 and GLM-5.1 are open-weight large language models from Z.ai (formerly Zhipu AI), the first two releases of the GLM-5 line that succeeded the GLM-4 generation. GLM-5, announced February 11–12, 2026, scaled the GLM mixture-of-experts (MoE) backbone to 744 billion total parameters (40 billion active) and was aimed at complex systems engineering and long-horizon agentic tasks. GLM-5.1, released April 7, 2026, focused on sustaining agentic performance over much longer horizons — Z.ai reported it could work independently for up to eight hours in a single run. Both are distributed under the MIT License. The line continued with GLM-5.2 in June 2026.

GLM-5

Z.ai launched GLM-5 targeting complex systems engineering and long-horizon agentic tasks, framing the release as a shift "from vibe coding to agentic engineering." Compared with GLM-4.5, GLM-5 scaled from 355B parameters (32B active) to 744B (40B active) and increased pre-training data from 23T to 28.5T tokens. It integrated DeepSeek Sparse Attention (DSA) to reduce deployment cost while preserving long-context capacity, and was trained with the company's open-sourced asynchronous reinforcement-learning infrastructure, "slime" (Source: z.ai; Source: github.com).

Z.ai positioned GLM-5 against Claude Opus 4.5 on code-logic density and systems-engineering capability, and reported it as achieving best-in-class performance among open-source models on reasoning, coding, and agentic tasks. Selected self-reported results (thinking mode):

(†Verified Terminal-Bench 2.0 variant. Source: z.ai.) On Vending Bench 2, an Andon Labs benchmark requiring a model to run a simulated vending-machine business over a one-year horizon, GLM-5 finished with a final account balance of $4,432, which Z.ai reported as first among open-source models and approaching Claude Opus 4.5. Z.ai also emphasized an "Office" capability — turning text or source materials directly into .docx, .pdf, and .xlsx deliverables such as reports and spreadsheets — exposed through an Agent mode in the Z.ai application (Source: z.ai).

GLM-5 weights were released under the MIT License on Hugging Face and ModelScope, and the model was made available on api.z.ai, BigModel.cn, and the free Z.ai chat interface, with compatibility with Claude Code and the company's OpenClaw assistant framework. Z.ai reported support for deployment on non-Nvidia chips including Huawei Ascend, Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame, and Hygon through kernel optimization and quantization (Source: z.ai). OpenRouter listed GLM-5 at $0.60 per million input tokens (Source: openrouter.ai). Coverage framed the release in the context of Z.ai's status as a publicly listed company, describing GLM-5 as a frontier model shipped by "China's first public AI company" (Source: medium.com). Two efficiency-oriented variants followed: GLM-5-Turbo (March 15, 2026) for high-throughput agent workloads and GLM-5V-Turbo (April 1, 2026) for vision-based coding and GUI agent tasks (Source: docs.z.ai).

GLM-5.1

GLM-5.1, released April 7, 2026, was described by Z.ai as its next-generation flagship for agentic engineering, with stronger coding than GLM-5 and a focus on staying effective over long task horizons rather than plateauing early. Z.ai reported it could work independently for up to eight hours in a single run and aligned its overall capability with Claude Opus 4.6, built using multi-turn supervised fine-tuning, reinforcement learning, and a process-quality evaluation framework (Source: docs.z.ai; Source: z.ai). Selected self-reported results:

(Source: z.ai.) Z.ai reported GLM-5.1 reaching state-of-the-art on SWE-Bench Pro among the compared models. To illustrate long-horizon behavior it described three open-ended runs: optimizing a vector database (VectorDBBench) over 600+ iterations and 6,000+ tool calls to reach 21,500 queries per second, roughly six times the best prior single-session result; optimizing GPU kernels (KernelBench Level 3) to a 3.6× geometric-mean speedup, behind Claude Opus 4.6's 4.2×; and building a Linux-style desktop web application over an eight-hour run that progressively added a file browser, terminal, text editor, and other components. Z.ai presented these as evidence that GLM-5.1 extends the "productive horizon" over which additional runtime continues to yield improvement (Source: z.ai). GLM-5.1 weights were released under the MIT License on Hugging Face and ModelScope, with availability on api.z.ai, BigModel.cn, and compatibility with Claude Code and OpenClaw; it was also published on Nvidia's NIM catalog in April 2026 (Source: z.ai; Source: build.nvidia.com).

Training and architecture

The GLM-5 line carried forward the mixture-of-experts design of GLM-4.5 at larger scale (744B total / 40B active) and added DeepSeek Sparse Attention to control long-context inference cost. Post-training for both GLM-5 and GLM-5.1 used the "slime" asynchronous RL infrastructure, which decouples rollout generation from training to keep training hardware saturated during long agentic rollouts. GLM-5.1's training emphasized sustained multi-round optimization — revisiting reasoning and revising strategy across hundreds of rounds and thousands of tool calls — rather than first-pass accuracy alone. The subsequent GLM-5.2 release added an IndexShare technique to extend the context window to one million tokens; see GLM-5.2 (Source: z.ai; Source: z.ai). After the June 2026 US export-control restrictions on Anthropic's Mythos-class models, security researchers said GLM-5.2 can match the latest US models at finding software security bugs, though it still trails Anthropic's and OpenAI's systems on other tasks (Source: wsj.com). See Export Controls (AI).

Availability and pricing

Both models are open-weight under the MIT License, available on Hugging Face, ModelScope, the Z.ai API, BigModel.cn, OpenRouter, and the free Z.ai chat interface, with a subscription "GLM Coding Plan" and the ZCode agentic development environment. As the most capable model at its release, GLM-5.1 consumed Coding-Plan quota at 3× during peak hours and 2× off-peak, with an off-peak 1× promotion through the end of April 2026 (Source: z.ai; Source: docs.z.ai).

Safety and evaluations

Z.ai did not publish a frontier-style system card or safety case for GLM-5 or GLM-5.1. Several benchmark runs (CyberGym, KernelBench audits) were conducted or audited with third-party harnesses and competing models, but no independent UK or US AISI evaluation of either model was available. The reported CyberGym results note that some competing models occasionally refused tasks they flagged as security risks, which Z.ai said may lower those models' scores (Source: z.ai).

Reception

GLM-5 drew coverage framing Z.ai as a publicly listed frontier-model developer and comparing GLM-5 to the Claude Opus series on long-horizon and systems-engineering tasks (Source: medium.com). GLM-5.1 was discussed in developer communities as roughly comparable to Claude Opus 4.6 on several benchmarks, an assessment noted as notable for an openly licensed model (Source: reddit.com).

The GLM-5 technical report frames the release as transitioning "the paradigm of vibe coding to agentic engineering." Its technical contributions are DSA, adopted "to significantly reduce training and inference costs while maintaining long-context fidelity"; an asynchronous reinforcement-learning infrastructure that "drastically improves post-training efficiency by decoupling generation from training"; and novel asynchronous agent RL algorithms "enabling the model to learn from complex, long-horizon interactions." The last two target the long-horizon gap that UK AISI's cyber-range results and Mowshowitz's assessment both identify as where open models trail furthest.

Relationships

• developer: Zhipu AI
• instance-of: Open-Weight Frontier Models — MIT-licensed open weights.
• supersedes: GLM-4 generation (GLM-4 through GLM-4.7) — the GLM-4 generation.
• superseded by: GLM-5.2 — GLM-5.2.
• related: DeepSeek V4 Pro / V4 Flash, Kimi K2, Qwen3, Claude Opus 4.5, Claude Opus 4.6, Gemini 3 / Gemini 3 Pro, US-China AI Competition: Different Races, Different Metrics, Fast-Follow Problem
• depends-on: Scaling Laws

Sources

Primary: Z.ai blog posts for GLM-5 (z.ai) and GLM-5.1 (z.ai); Z.ai developer release notes (docs.z.ai); GitHub (github.com); Nvidia NIM model card (build.nvidia.com); OpenRouter (openrouter.ai). Secondary: Médialabonne/Medium GLM-5 write-up, Reddit developer discussion. The GLM-5 technical report is arXiv 2602.15763. Confidence medium: release dates, scale, licensing, and headline benchmarks are corroborated across the official posts and release notes, but benchmarks are self-reported and not independently replicated.