@zhivex-ai/meta

Meta Model API adapter for Zhivex AI SDK. muse-spark-1.3 is the recommended direct Meta model; the official muse-spark-1.3-contributor and muse-spark-1.2-contributor variants and older IDs such as muse-spark-1.1 remain usable when they are available to your Meta project.

Install

bun add @zhivex-ai/core @zhivex-ai/meta

Usage

import { generateText } from "@zhivex-ai/core";
import { createMeta } from "@zhivex-ai/meta";

const meta = createMeta({
  apiKey: process.env.MODEL_API_KEY
});

const result = await generateText({
  model: meta("muse-spark-1.3"),
  prompt: "Explain tool calling in one sentence."
});

console.log(result.text);

The adapter defaults to https://api.meta.ai/v1 and reads MODEL_API_KEY when no key is passed explicitly.

API modes and provider features

Chat Completions is used for ordinary text, vision, callable tools, and structured output. Set providerOptions.apiMode to "responses", or use files or Meta hosted tools, to use the Responses API:

import { generateText } from "@zhivex-ai/core";
import { createMeta, metaToolSearchTool, metaWebSearchTool } from "@zhivex-ai/meta";

const meta = createMeta({ apiKey: process.env.MODEL_API_KEY });

const result = await generateText({
  model: meta("muse-spark-1.3"),
  prompt: "Find the current primary source and summarize it.",
  tools: {
    webSearch: metaWebSearchTool(),
    toolSearch: metaToolSearchTool({
      execution: "client",
      description: "Find a tool by name.",
      parameters: { type: "object", properties: { query: { type: "string" } } }
    })
  },
  providerOptions: {
    prompt_cache_key: "research-prefix-v1",
    prompt_cache_retention: "24h"
  }
});

Responses streaming supports text deltas, fragmented function-call arguments, and continuation through previous_response_id. Retryable HTTP statuses (408, 429, and 5xx) are retried when maxRetries is configured, before a JSON body or SSE stream is consumed.

Chat Completions and Responses preserve Meta's reported reasoning count in usage.reasoningTokens, including Responses streaming and grounded generation. Reasoning tokens are already included in usage.outputTokens; do not add them again to output or total tokens. A reported zero remains 0, while an unreported reasoning count remains undefined.

Meta Model API accepts only toolChoice: "auto" (which is also the default). The adapter rejects "none", "required", and named-tool choices before sending a request.

The shared Zhivex audio part is supported for MP3 and WAV input. Chat Completions sends inline base64 audio; Responses additionally accepts base64 data URLs and uploaded Meta file IDs. Audio output is not supported.

Meta Model API has a native computer tool on Responses. The current Zhivex Meta adapter does not yet map native computer_call and computer_call_output items, so its native computerUse capability remains false. Use runComputerUse() for the SDK's portable function-and-screenshot path. The native probe demonstrates the direct Meta protocol against an isolated page.

Repository and full documentation:

• https://github.com/Zhivex/zhivex-ai-sdk
• https://developer.meta.com/ai/models/muse-spark/
• https://dev.meta.ai/docs/protocols/responses
• https://dev.meta.ai/docs/video-understanding
• https://dev.meta.ai/docs/computer-use

Muse Spark 1.3

The SDK catalog includes muse-spark-1.3 and muse-spark-1.3-contributor. Pass either exact ID to languageModel() to use the existing Chat or Responses adapter, including structured output, tools, streaming, and reasoning. Meta's announcement confirms max reasoning availability for Standard 1.3. Pricing is omitted from the catalog; no 1.2 price is inherited. Existing 1.2 integrations and smoke defaults remain available.

See Meta's release announcement.

File uploads preserve the exact bytes of a supplied Uint8Array or Buffer view, including slices with nonzero offsets.