Zhipu GLM Agent
Type: full-code · Vendor: Zhipu AI (Z.ai) · Language: API · License: proprietary (hosted API) / MIT (GLM open weights since July 2025) · Status: active · Status in practice: emerging
Tsinghua-spinoff Zhipu (rebranded internationally as Z.ai in July 2025) ships the GLM model line through the BigModel platform, with function calling, web search, knowledge-base retrieval and JSON output all available as tool/response options on the same chat-completion API; AutoGLM is the consumer agent built on top.
Description. Zhipu (Z.ai) is one of China's three biggest LLM providers and IPO'd on the Hong Kong Stock Exchange on 8 January 2026. Its GLM model line has been MIT-open-source since July 2025 (GLM-4.5 onward), with the latest GLM-5/5.1 and GLM-5V-Turbo positioned for agentic engineering, coding and long-chain execution. The BigModel chat-completion API supports tools: function calling, knowledge-base retrieval and web search are listed together as tool options, and structured output is selectable via response_format json_object. AutoGLM is the consumer agent application that uses voice commands to complete tasks on a phone and can self-execute operations exceeding 50 steps across applications.
Agent loop shape. Two surfaces. (1) BigModel chat-completion API: standard tools-loop where tools[] can include functions, knowledge-base retrieval, or web search; the model emits tool_calls, the caller executes, results return as tool messages; response_format json_object pins shape. (2) AutoGLM: hosted consumer agent that takes voice commands, plans multi-step actions, and self-executes >50-step workflows across smartphone apps.
Primary use cases
- function-calling agents over GLM hosted API
- RAG via knowledge-base retrieval and web search as built-in tools
- JSON-structured extraction via response_format
- consumer voice agent (AutoGLM) on smartphones
Key concepts
- GLM-4.6 / GLM-5 / GLM-5.1 (docs) — Flagship Zhipu models; GLM-4.6 raises context to 200K, GLM-5/5.1 positioned as base models for agentic engineering with open-source SOTA-level coding.
- Tools parameter (function/retrieval/web search) → tool-use (docs) — Single tools parameter accepts function calls, knowledge-base retrieval, and web search; model emits a JSON-typed call the caller handles.
- response_format json_object → structured-output (docs) — Forces the model to return structured JSON; documented as suitable for extraction and structured API responses.
- AutoGLM → mobile-ui-agent (docs) — Consumer agent application that uses voice commands to complete tasks; can self-execute operations exceeding 50 steps across smartphone apps.
- GLM-5-Turbo for agent harnesses (docs) — Trained at the training level for core agent capabilities — tool calling, instruction following, long-chain execution — targeted at agent runtimes ('Claw').
Patterns this full-code implements —
- ★★Agentic RAG
Web search and knowledge-base retrieval are listed together with function calling as values the tools[] parameter accepts — retrieval is model-callable, not a separate API. Knowledge-base retrieval i…
- ★★Structured Output
response_format json_object is a top-level toggle that constrains the model to valid JSON; BigModel also documents JSON-Schema-based structured output on GLM-4.6 / GLM-4.7 with explicit Schema design…
- ★★Tool Use
Function calling is a documented chat-completion capability; the model emits a function name + JSON arguments for the caller to execute.
- ★Mobile UI Agent
AutoGLM is positioned as a voice-driven smartphone agent that self-executes >50-step cross-application workflows; this is a separate consumer surface from the BigModel API.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified