@zhivex-ai/zai

Native Z.ai GLM adapter for Zhivex AI SDK.

Install

bun add @zhivex-ai/zai @zhivex-ai/core zod

GLM-5.3 Flash

glm-5.3-flash is available through the general Model API and GLM Coding Plan. It has a 1M-token context window, native image understanding, required thinking, streamed tool calls, and the same text parameter contract as GLM-5.3. The general pay-as-you-go endpoint remains the provider default:

import { generateText } from "@zhivex-ai/core";
import { createZAI } from "@zhivex-ai/zai";

const zai = createZAI({ apiKey: process.env.ZAI_API_KEY });

const result = await generateText({
  model: zai("glm-5.3-flash"),
  messages: [
    {
      role: "user",
      parts: [
        { type: "text", text: "Review this architecture diagram." },
        {
          type: "image",
          image: "https://example.com/architecture.png",
          mediaType: "image/png"
        }
      ]
    }
  ],
  reasoning: { effort: "max" }
});

console.log(result.text);

The default base URL is https://api.z.ai/api/paas/v4. ZAI_BASE_URL or the baseURL option can override it for compatible gateways. Image parts preserve their message order and accept public HTTP(S) URLs, image Data URLs, or raw base64 normalized to a Data URL. Multiple images are supported; audio, video, and generic file parts remain outside this model contract.

General API and Coding Plan

Use the default general endpoint for pay-as-you-go glm-5.3-flash, glm-5.3, or glm-5.2 calls. Select the Coding Plan endpoint explicitly when using a Coding Plan credential:

const coding = createZAI({
  apiKey: process.env.ZAI_API_KEY,
  endpoint: "coding"
});

const result = await generateText({
  model: coding("glm-5.3-flash"),
  prompt: "Review this authorization boundary.",
  reasoning: { effort: "max" }
});

Coding Plan is limited by Z.ai's usage policy to supported coding and agent tools. Use the general endpoint when exact pay-as-you-go routing or model behavior matters.

Contract

• Text generation and SSE streaming through Chat Completions.
• Local function-tool loops with automatic tool selection. Forced, required, none, hosted tools, and provider-advertised parallel calls are not exposed because only tool_choice: "auto" is currently documented.
• reasoning_content is preserved as Z.ai provider-data and replayed unchanged across tool turns. The adapter sends thinking.clear_thinking: false when tool state must be preserved.
• GLM-5.3 and GLM-5.3 Flash require thinking and accept shared efforts low, high, and max. none, disabled, and unsupported aliases fail before network I/O. Flash defaults thinking.clear_thinking to false, matching the model-specific recommendation.
• GLM-5.2 maps none/minimal to disabled thinking, low/medium to high, and xhigh to max.
• GLM-5.3 Flash exposes image understanding through ordered image_url content blocks. Other Z.ai models remain text-only until their individual media contracts are modeled.
• Structured output uses Z.ai JSON-object mode plus a schema prompt, followed by local Zod validation. Z.ai does not document strict native JSON Schema enforcement.
• Cached input usage and documented provider finish reasons are normalized into the shared contract.
• The adapter does not claim embeddings, audio, realtime, hosted web search, or remote MCP. Vision is exposed only for GLM-5.3 Flash.

Use providerOptions for documented Z.ai fields such as thinking, reasoning_effort, top_p, do_sample, stop, request_id, and user_id. Do not put private information in provider user identifiers.

Live smoke

General GLM-5.3 Flash smoke:

ZAI_API_KEY=... \
ZAI_INTEGRATION_MODEL=glm-5.3-flash \
ZAI_EXTENDED_INTEGRATION=1 \
bun run test:integration:zai

Add ZAI_INTEGRATION_IMAGE_URL=https://... and optionally ZAI_INTEGRATION_IMAGE_MEDIA_TYPE=image/png to include the authenticated vision smoke.

Coding Plan GLM-5.3 Flash smoke:

ZAI_API_KEY=... \
ZAI_ENDPOINT=coding \
ZAI_INTEGRATION_MODEL=glm-5.3-flash \
ZAI_EXTENDED_INTEGRATION=1 \
bun run test:integration:zai

A skipped smoke or offline fixture test is not authenticated provider certification.