Models

Meta ships five model families. Muse Spark, Muse Image, Muse Voice Transcribe, and Segment Anything Model (SAM) are hosted on Meta Model API and called over the endpoints in these docs; Muse Glimmer is open-weight and runs on your own hardware. This page summarizes each and points you to where to go next.

Muse Spark

Muse Spark is Meta's model for agentic and coding work — multi-step tool loops, software engineering assistants, and long-context reasoning. Use it for chat completion, image understanding, video and audio understanding, tool calling, structured output, and search grounding.

Muse Spark comes in three versions, each sharing the same modalities and context window, and differing only by capability:

• Muse Spark 1.3 (muse-spark-1.3): the latest version, tuned for agentic workflows (multi-step tool, browser, and long-horizon tasks) with improved coding over 1.2. Supports all reasoning effort levels, including the "max" level for extended reasoning (available on Standard tier only). Recommended for new work.
• Muse Spark 1.2 (muse-spark-1.2): the previous version.
• Muse Spark 1.1 (muse-spark-1.1): the original version.

muse-spark-1.3 is the default model in the code examples throughout these docs.

Available Muse Spark models

* Audio understanding in Muse Spark 1.3 is currently not fully supported, and response quality for requests including audio content may be degraded. For audio, use Muse Spark 1.2, or Muse Voice Transcribe for dedicated speech-to-text.

Tiers

Every version is offered on the Standard tier (standard pricing; your data is never used for training). Versions 1.3 and 1.2 also offer a discounted Contributor variant (muse-spark-1.3-contributor, muse-spark-1.2-contributor) that trades a lower price for permission to train on your prompts and completions — room to prototype and scale experiments where that's acceptable. See Pricing and rate limits for per-tier rates.

Muse Image

Muse Image is a separate model family that outputs images rather than text. Send a text prompt (and optional reference images) and get an image back, or send an image with an instruction to edit it. One model handles both generation and editing. It's agentic, so it can search the web for visual references and current facts and run code to build layouts before it renders.

Reach Muse Image two ways: the Responses API for conversational, multi-turn editing, or the single-shot /v1/images/generations and /v1/images/edits endpoints for one-off calls. See the Image generation guide to get started.

Muse Voice Transcribe

Muse Voice Transcribe is Meta's speech-to-text model on Meta Model API. Use it to transcribe live audio streams or supported audio files.

Muse Voice Transcribe is built for real-time speech products: voice agents, meeting and call intelligence, live transcription, dictation, captioning, and high-volume transcription. It supports streaming speaker diarization, native endpointing and voice activity detection, contextual and keyword biasing, 25 evaluated languages with code-switching, and turn-level timestamps.

It returns transcript text. It does not synthesize speech or provide a speech-to-speech conversation API. It does not provide word-level timestamps, sound event detection, or emotion detection.

Reach Muse Voice Transcribe two ways: the realtime WebSocket endpoint (wss://api.meta.ai/v1/asr/realtime) for live audio, or the file endpoint (POST /v1/asr/transcribe) for a recording you already have.

Start with the Muse Voice Transcribe guide. See Pricing and rate limits for audio pricing.

Segment Anything Model

SAM 3.1 is Meta's segmentation model on Meta Model API. Give it a short noun phrase such as "yellow school bus" with an image or video, and it returns every match as a box and a pixel-accurate mask. One model handles both stills and clips, and in video it follows each object across frames.

Reach SAM 3.1 over the Responses API (POST /v1/responses), and upload clips through the Files API when you segment them more than once. See the Media segmentation guide to get started.

Muse Glimmer

Muse Glimmer is Meta's open-weight multimodal model, distilled from Muse Spark and built to run on your own hardware. Unlike Muse Spark, Muse Image, Muse Voice Transcribe, and SAM, you don't call it over Model API — you download the weights and serve it through a runtime such as vLLM, SGLang, llama.cpp, or ExecuTorch.

Muse Glimmer has its own documentation section covering how to get, prompt, deploy, and customize the model:

• Muse Glimmer overview: variants, architecture, license, and launch partners.
• Get the model: download the weights and verify your setup.
• Run inference: pick a runtime and serve it locally.

Model availability

Each model is served on a specific set of Meta Model API endpoints. Find where to call the one you want:

Muse Glimmer is self-hosted and isn't served on any Meta Model API endpoint.

List models via the API

Query the catalog programmatically when you need to check what's enabled for your team. This returns the API-hosted models; self-hosted Muse Glimmer isn't included.

python
import os
​
from openai import OpenAI
​
client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)
​
response = client.models.list()
​
print(response.model_dump_json(indent=2))

python
import json
import os
​
import requests
​
response = requests.get(
    "https://api.meta.ai/v1/models",
    headers={"Authorization": f"Bearer {os.environ['MODEL_API_KEY']}"},
)
response.raise_for_status()
print(json.dumps(response.json(), indent=2))

shell
curl -X GET "https://api.meta.ai/v1/models" \
  -H "Authorization: Bearer $MODEL_API_KEY"

The response returns models sorted newest-first by creation time, with the model ID as a stable tiebreaker. Each object includes a created field — the Unix timestamp in seconds for when the model was added to the registry. GET /v1/models/{model} returns the same created value for a single model.

Next steps

• Get started: make your first call with Muse Spark
• Pricing and rate limits: see rates and retry guidance
• Chat completion: start generating with the core conversational endpoint