Robots Atlas>ROBOTS ATLAS
GLM-5
AI Models›GLM

GLM-5

5 · Family: GLM
Z.ai's flagship foundation model designed for Agentic Engineering. 744B/40B MoE, 28.5T tokens, DeepSeek Sparse Attention. SOTA open-weight on SWE-bench Verified 77.8 - coding comparable to Claude Opus 4.5.
✓ Active✓ Public access⚖ Open weights★ FeaturedLLMReasoning modelTool-using model📁 GLM
Context window
200K
tokens
Parameters
744B total / 40B activated (MoE)
parameters
Max output
128,000
tokens
Release date
11 February 2026
Access:APIDownloadHostedDeployment:☁ Cloud💻 Local

Overview

GLM-5 is Z.ai's new-generation flagship foundation model (formerly Zhipu AI, headquartered in Beijing), officially announced in February 2026. The model is designed for Agentic Engineering - multi-step task chains, long-horizon planning, code refactoring, and autonomous execution of complex engineering projects. Z.ai positions GLM-5 as SOTA open-weight in coding and agentics - its usability in real programming scenarios approaches Claude Opus 4.5.

Architecture: parameter scale increased from 355B (32B activated) in GLM-4.7 to 744B total / 40B activated (MoE - Mixture of Experts). Pre-training data expanded from 23T to 28.5T tokens. The model is the first in the GLM family to integrate DeepSeek Sparse Attention - a sparse attention mechanism maintaining lossless long-text quality with significantly lower deployment cost and better token efficiency. Post-training stack: a new Slime framework for asynchronous reinforcement learning - supports larger model scales and more complex RL tasks, including an asynchronous agent RL algorithm learning from long-horizon interactions.

Technical specification: 200K token context window, maximum output 128K tokens. Modalities: text-in / text-out (no native image, video, or audio support - those are in separate GLM-V/GLM-Omni family models). Built-in capabilities: thinking mode (switchable, enabled by default), streaming output, function calling (tool use), context caching (KV cache), structured output (JSON mode).

Benchmark results (SOTA among open-weight): SWE-bench Verified: 77.8 (leading open-model score, surpasses Gemini 3.0 Pro), Terminal Bench 2.0: 56.2, top results on BrowseComp (web-scale retrieval), MCP-Atlas (tool invocation and multi-step task execution), and τ²-Bench (multi-tool planning and orchestration). Capable of autonomous long-horizon planning, backend refactoring, and debugging with minimal human intervention.

The model is available via the Z.AI Open Platform (api.z.ai, OpenAI-compatible format at api.z.ai/api/paas/v4/) and the z.ai chat. Access to GLM-5 (API name: glm-5) requires a GLM Coding Plan Pro or Max subscription - monthly plans for world-class model access, compatible with Claude Code, Open Code, and other tools. Official SDKs: zai-sdk (Python), ai.z.openapi:zai-sdk (Java), plus OpenAI Python SDK support. Weights are released as open-weight - GLM-5 continues the openness policy of the GLM family, contrasting with Chinese competitors (Qwen3-Max, DeepSeek V4 - both proprietary).

Classification
LLMReasoning modelTool-using model
Family: GLM
Access & deployment
APIDownloadHosted
CloudLocal
Weights: Open weights
Key parameters
📏 Context: 200K
🧩 Parameters: 744B total / 40B activated (MoE)
✓ Tools · ✓ Fine-tuning
📥 Input: text, structured data, documents

Technical specification

Context window
200K
tokens
Parameters
744B total / 40B activated (MoE)
parameters
Max output tokens
128,000
tokens per response
Knowledge cutoff
1 Dec 2025
Knowledge boundary
License
Open weights (MIT-style, patrz repozytorium Z.ai / GitHub THUDM/ZhipuAI - GLM Coding Plan Pro/Max wymagany do dostępu przez api.z.ai)
Hardware requirements
Open-weight model - can be self-hosted on multi-GPU infrastructure (recommended min. 8x H100/H200 for the 744B MoE, ~120GB VRAM at 40B activated with INT8 quantisation). Alibaba/Z.ai also serve the model commercially via api.z.ai and integrations with ModelScope/HuggingFace.
Features:✓ Tool use✓ Fine-tuning
Modalities
⬇ Input
textstructured_datadocuments
⬆ Output
textcodestructured_data

Capabilities and applications

Native model capabilities
Advanced reasoning
The ability to perform multi-step, structured reasoning: analysing problems, planning steps, and drawing conclusions from hypotheses. Reasoning-first models (e.g. GPT-5.1 Thinking) dedicate a portion of inference to chains of thought before responding.
Category: reasoning
Extended thinking mode
A reasoning-model variant with a larger inference budget: more thinking cycles, higher answer precision at the cost of response time. Choice between 'standard' and 'extended' thinking is left to the user (e.g. the selector in GPT-5.2 Pro).
Category: reasoning
Adaptive reasoning effort
The model decides how much 'thinking' to allocate to a given query: simple questions are answered quickly, complex problems receive more inference cycles. A GPT-5.1 feature (both Instant and Thinking) that shortens time on easy tasks and extends it for hard ones.
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Mathematical reasoning
The model's ability to solve mathematical tasks requiring multi-step reasoning — equations, proofs, combinatorics, geometry, calculus and competition-level problems.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Agentic coding
Multi-hour, multi-step programming tasks performed autonomously by the model: cloning a repository, running tests, iterating on fixes, integrating with CLI tools. Characteristic of Codex variants (GPT-5.1-Codex-Mini, Codex-Max).
Category: coding
Agentic capability
The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.
Category: planning
Tool use
The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.
Category: planning
Long context
Support for large context windows — tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Prompt caching
Cost-performance optimisation: repeated prompt fragments (e.g. system prompt, long documentation) are cached server-side and cheaper in subsequent calls. Significantly reduces cost for applications with long contexts.
Category: other
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language

Benchmark results

5 benchmarks
SWE-bench Verified
resolved rate · Leading open-model score, agentic real-world GitHub issue fixing
77.8%
📅 11 Feb 2026📄 Z.ai GLM-5 blog (docs.z.ai/guides/llm/glm-5)
Terminal Bench 2.0
success rate · Terminal-based multi-step tasks, surpasses Gemini 3.0 Pro
56.2%
📅 11 Feb 2026📄 Z.ai GLM-5 blog (docs.z.ai/guides/llm/glm-5)
BrowseComp
task completion · Web-scale retrieval and information synthesis
SOTA open-model%
📅 11 Feb 2026📄 Z.ai GLM-5 blog
MCP-Atlas
task completion · Tool invocation and multi-step task execution via MCP
SOTA open-model%
📅 11 Feb 2026📄 Z.ai GLM-5 blog
τ²-Bench (Tau-Bench 2)
success rate · Complex multi-tool planning and orchestration
SOTA open-model%
📅 11 Feb 2026📄 Z.ai GLM-5 blog

Pricing

Technical architecture

Deployment and security

🔒 Security / Enterprise
✓ Verified enterprise information

The model is hosted on Z.ai's infrastructure in China. Due to its PRC headquarters, the model is subject to Cyberspace Administration of China AI regulations. Customers requiring data residency outside China have an alternative: self-hosting the open-weights on their own infrastructure (private cloud, on-premise, air-gapped). Open-source SDK code allows full integration audit.

↗ Security documentation

Sources and related pages