BeamWeaver Z.ai

BeamWeaver includes a Z.ai provider under BeamWeaver.ZAI for GLM-5.3, GLM-5.3-Flash, and GLM-5.2 Chat Completions.

Surface

• BeamWeaver.ZAI.ChatModel implements BeamWeaver.Core.ChatModel.

• BeamWeaver.ZAI.Client calls https://api.z.ai/api/paas/v4/chat/completions by default.

• Model initialization is strict: use zai:glm-5.3, zai:glm-5.3-flash, or zai:glm-5.2. Bare glm-* identifiers and other zai:* models are rejected before transport.

• Runtime config uses config :beam_weaver, :zai; runtime.exs reads ZAI_API_KEY plus optional ZAI_BASE_URL or ZAI_API_URL.

• Standard function tools are rendered as OpenAI-compatible Chat Completions tools. Z.ai tool_stream: true is supported for streaming tool-call argument chunks and requires stream: true.

• Structured output uses JSON object mode: response_format: %{type: "json_object"}. BeamWeaver schema requests are mapped to JSON object mode, the schema is injected as provider-visible instructions, and the response is parsed and validated locally. JSON Schema request mode is not enabled for this provider.

• Streaming reconstructs text, reasoning content, streamed tool-call chunks, final usage chunks, and finish_reason: "length" truncation.

• Usage metadata tracks prompt, completion, total, cached input, and reasoning output tokens. Cost metadata uses the selected profile's checked-in prices and does not double-bill reasoning tokens.

• Token counting uses BeamWeaver's approximate fallback.

Usage

config :beam_weaver,
  zai: [
    api_key: System.fetch_env!("ZAI_API_KEY")
  ]

{:ok, model} =
  BeamWeaver.Models.init_chat_model("zai:glm-5.3",
    reasoning_effort: :high,
    thinking: %{type: :enabled},
    max_output_tokens: 1_024
  )

{:ok, message} =
  BeamWeaver.Core.ChatModel.invoke(model, [
    BeamWeaver.Core.Message.user("Reply with a concise plan.")
  ])

Stream with usage and tool-call argument chunks:

{:ok, message} =
  BeamWeaver.ZAI.ChatModel.stream_response(
    model,
    [BeamWeaver.Core.Message.user("Call get_weather for Tokyo.")],
    tools: [weather_tool],
    tool_choice: "auto",
    tool_stream: true
  )

Profiles

All three profiles are checked in with:

• 1,000,000 input tokens

• 131,072 maximum output tokens

• text input and output

• reasoning output

• function tools

• JSON object mode

• streaming

• usage metadata

• Chat Completions API only

GLM-5.3 and GLM-5.3-Flash always use thinking and accept low, high, and max reasoning effort; max is the default. GLM-5.2 retains the broader compatibility effort ladder and supports enabled or disabled thinking.

GLM-5.3-Flash additionally accepts native image, video, and PDF inputs and attachment content. The other two profiles are text-input models.

Current cost metadata per one million tokens:

The GLM-5.3-Flash promotional rates are recorded through 2026-09-09T24:00:00+08:00; regular rates remain in the profile so accounting can switch at the explicit boundary rather than guessing.

Unsupported Z.ai Surfaces

• Z.ai models outside the three explicit IDs above are intentionally not routed.

• JSON Schema request mode is not enabled until live or documented support is clear for this endpoint.

• Built-in tools and non-chat APIs are not exposed in BeamWeaver. Media input is limited to the GLM-5.3-Flash profile.