Zhipu (GLM)

TL;DR

ONGRID_ZHIPU_API_KEY=...
ONGRID_ZHIPU_MODEL=glm-4.7                  # default
ONGRID_ZHIPU_BASE_URL=                      # optional; defaults to open.bigmodel.cn

Provider id: zhipu. SDK adapter: OpenAI-compatible Chat Completions against Zhipu's bigmodel.cn endpoint.

Zhipu's GLM family is the most-used Chinese-region provider for Ongrid installs. The bigmodel.cn endpoint is OpenAI-compatible at the wire level, so the same eino-openai adapter handles it.

Env vars

Var	Default	Notes
ONGRID_ZHIPU_API_KEY	—	Empty = provider dropped
ONGRID_ZHIPU_MODEL	glm-4.7	Default model
ONGRID_ZHIPU_BASE_URL	https://open.bigmodel.cn/api/paas/v4	Override for self-hosted or VPC endpoints
ONGRID_ZHIPU_MODELS	glm-5.1,glm-5,glm-4.7,glm-4.7-flash	Catalog list

Default catalog

The out-of-box picker shows:

• glm-5.1 — frontier; use for the hardest RCA passes.
• glm-5 — general default for the GLM family.
• glm-4.7 — the catalog default; balanced cost/quality.
• glm-4.7-flash — cheap; recommended for Pass-2 extraction.

The default model glm-4.7 is the conservative pick — GLM-4.7 has been battle-tested against the investigator persona; 5.x is still being benchmarked for the tool-use loop.

BaseURL — China-based marker

Zhipu is China-based: the default https://open.bigmodel.cn/api/paas/v4 endpoint resolves only from networks that can reach mainland CN infrastructure. Most non-CN clouds need either a VPC peering or a relay.

The Settings UI tags the BaseURL field as "China-based" so admins outside CN know they need a relay. The same tag exists for Kimi (Moonshot) — see Kimi.

For self-hosted variants of GLM (glm-zero open-weights), use the Custom provider slot instead — Zhipu's slot is wired for bigmodel.cn's API shape specifically.

Making Zhipu the default

ONGRID_LLM_DEFAULT_PROVIDER=zhipu

The LLM-resolver-bug memo: when default_provider is unset but multiple providers are configured, the resolver picks the first-sorted provider id and uses ITS default model. A subtle mismatch follows: if you set ONGRID_OPENAI_MODEL=glm-4.7 thinking the model wins, the router still routes to openai and sends glm-4.7 to OpenAI's endpoint. Fix: set default_provider explicitly.

Quirks

• OpenAI-compatible at the wire — tool calls, streaming, system messages all use OpenAI's shapes. No special adapter.
• Prompt sensitivity — GLM follows user-message constraints more strictly than system-message ones. The investigator persona's 10-tool-call budget is in the user message specifically because of this; system-message budgets were ignored in v0.7.51 - v0.7.55.
• Locale — GLM defaults to Chinese responses unless the prompt explicitly says otherwise. The LANGUAGE: ... directive Ongrid appends to the investigator prompt was originally introduced for GLM. See [feedback_ai_output_locale].

See also

• Models overview.
• Routing — WithProvider("zhipu").
• RCA — why the budget directive lives in the user message and how the locale flag works.