Muse Spark Open Documentation

Download weights and deploy the model

Get official weights from Meta’s Muse Spark Open collection on Hugging Face. The announced ecosystem includes SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth. Choose a framework supported by your hardware and verify its model-specific version, memory and quantization requirements. For a hosted evaluation, sign in to the playground and review the connected model and available controls. For API integration, create an account API key and follow the Chat Completions examples below.

• Muse Spark Open download
• Meta developer guide

1. Read the model card and license.

2. Choose weights, hardware and a runtime.

3. Evaluate your coding and agent tasks.

Prepare multimodal inputs and task context

The model accepts text, images, video and audio and returns text. Give it a concrete objective, relevant files or media, and a clear output format. For coding agents, include repository conventions and explicit tool permissions. The current playground supports text and compatible image/video URLs. Audio input and tool execution require a separate compatible deployment or integration. The published 1M context is a model capability, not the upload limit of this page.

Goal: review this repository change.
Context: the relevant files, tests and deployment constraints.
Output: correctness issues, supporting code references and a minimal repair plan.
Tools: request permission before applying changes.

Understand responses and conversation history

Read and verify model output before using it. In this playground, credits are reserved before inference and the response is released after billing settlement. Stop interrupts the request. Conversation context stays on this page; refreshing starts a new conversation. Copy answers you need before leaving. The official model supports low, medium, high and xhigh reasoning effort. For Muse Spark Open, use those values; the generic Max option is not supported by its model card. The service connection determines which controls are available.

Muse Spark Open API

Use our OpenAI-compatible Chat Completions API at POST /api/v1/chat/completions. Authenticate with Authorization: Bearer and select model meta/muse-spark-open. Responses use the standard choices and usage fields. The endpoint accepts text, image/video URLs, base64 WAV/MP3 audio and function-tool messages. Requests are limited to 4 MiB and 128 messages, with max_tokens from 1 to 16,384 (default 2,000). The model requests tool calls; your application executes them. Credits are reserved before inference and settled before output is delivered, including streaming output. Rejected upstream requests are refunded. Interrupted requests or missing final costs may retain the prepaid amount; check credit history or contact support.

POST https://themusesparkopen.com/api/v1/chat/completions
Authorization: Bearer $MUSE_SPARK_API_KEY
Content-Type: application/json
Model: meta/muse-spark-open

export MUSE_SPARK_API_KEY="YOUR_API_KEY"

const response = await fetch('https://themusesparkopen.com/api/v1/chat/completions', {
  method: 'POST',
  headers: {
    Authorization: `Bearer ${process.env.MUSE_SPARK_API_KEY}`,
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    model: 'meta/muse-spark-open',
    messages: [{ role: 'user', content: 'Hello!' }],
    max_tokens: 2000,
    reasoning: { effort: 'medium' },
  }),
});

const result = await response.json();
if (!response.ok) {
  throw new Error(result.error?.message || 'Request failed');
}
console.log(result.choices[0].message.content);

curl -N 'https://themusesparkopen.com/api/v1/chat/completions' \
  -H "Authorization: Bearer $MUSE_SPARK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta/muse-spark-open",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 2000,
    "reasoning": {"effort": "medium"},
    "stream": true
  }'

Work with the account assistant

The model is designed for multi-step planning, coding, tool invocation and recovery across extended tasks. An agent application must supply its tools, validate arguments and decide what may execute. Workflow snippets on the homepage illustrate a compatible deployed client; they are not the playground request schema. The account assistant uses the separately configured models and tools for this site. Media-generation tools use their own providers; Muse Spark Open itself produces text. Assistant history and playground conversations are separate.

Manage credits and choose a plan

Muse Spark Open costs $1.25 per 1M input tokens and $4.25 per 1M output tokens (USD). API and playground requests use account credits. Check Pricing for the current subscriptions and credit packs. The playground reserves credits before sending a request and settles against reported usage. Stopping, timeouts or missing final cost may retain a reservation; see your credit history and contact support for billing questions.

Resolve integration and account issues

Check your session, credit balance and the message displayed when a request fails. Use the supported input types and reasoning settings, and reduce oversized requests. If the model connection is unavailable, contact support@themusesparkopen.com. For model deployment, follow the official model card and framework documentation.