Skip to content

2026-08-14-glm-5.3.md

Latest commit

 

History

History
42 lines (35 loc) · 2.26 KB

2026-08-14-glm-5.3.md

File metadata and controls

42 lines (35 loc) · 2.26 KB
Raw

GLM-5.3 support and a unified GLM client

  • Date: 2026-08-14
  • Type: feature
  • Scope: glm5_3, glm5_1, glm5_2, auto_client, registry
  • PR: #169

中文版

AgentHub now supports Z.AI/Zhipu's glm-5.3 (announced at https://z.ai/blog/glm-5.3), and the GLM series clients merge into one, following the Gemini unification: glm5_1/ and glm5_2/ fold into a single glm5_3/ client per language, named for the newest generation. Outside the thinking parameter contract the 5.1 and 5.2 implementations differed only by the reasoning_effort mapping, so per-generation rules now live in one client.

Implementation notes

  • Docs-only implementation. The GLM-5.3 API was announced but not live at the time of this change, so there is no live capture; the implementation follows the official pre-launch documentation at docs.bigmodel.cn, snapshotted under llmsdk_docs/glm5_3/. glm-5.2 stays as the env-gated E2E model until the API is live.
  • Per the docs, GLM-5.3 keeps the GLM-5.2 wire format (1M context, 128K max output, streaming reasoning_content/content/tool_calls, tool_stream, tool_choice auto only) and changes only the thinking contract:
    • GLM-5.3 uses forced thinking and errors on thinking.type: "disabled", so ThinkingLevel.NONE degrades to the new light-reasoning low effort with thinking enabled.
    • GLM-5.3 accepts only low/high/max for reasoning_effort and errors on other values, so the client clamps: LOW → low, MEDIUM/HIGH → high, XHIGH → max. GLM-5.2 keeps the full pass-through vocabulary (the server maps it), and pre-5.2 models still send no reasoning_effort at all.
  • Routing folds to the glm-5 prefix, which every previously accepted GLM client type contains, so glm-5.1/glm-5.2 spellings keep working; the supported-types message lists glm-5.3 first. Official registry entries were repointed at glm-5.3; the OpenRouter-hosted z-ai/glm-5.2 entry was not changed.

Registry metadata

glm-5.3 is registered with text input/output and a 1,000,000-token context window. Pricing is omitted: the API is not yet live and Z.AI publishes no GLM-5.3 price at snapshot time.