Muse by Meta

Muse is Meta's family of AI models for developers. The reasoning models (Muse Wonder, Muse Spark, and Muse Ember), plus models for image generation, speech-to-text, and segmentation, are offered through Meta Model API. This pay-as-you-go inference API is served at https://api.meta.ai/v1. Meta also publishes two open-weight models, Muse Spark Open and Muse Glimmer. Developers use the family to build their own agents and applications. Meta sells its Muse Code coding agent separately.

Key Capabilities

• Reasoning models: Muse Wonder, Muse Spark, and Muse Ember accept text, image, video, audio, and PDF input, return text, and have a 1,048,576-token context window. They take the same request body, so moving a workload between them is a model ID change. Reasoning effort is set per request from "minimal" to "max".
• Agent features: Tool calling supports parallel and streamed calls and on-demand loading of large tool sets. Search grounding returns answers with inline citations, and computer use drives a desktop from screenshots. Structured output returns JSON matching a supplied schema. Prompt caching with token counting and file upload by ID are also available.
• Muse Image: Generates and edits images from a text prompt through /v1/images/generations and /v1/images/edits. Multi-turn refinement runs through the Responses API.
• Muse Voice Transcribe: Speech-to-text for streaming audio over a WebSocket and for files. It provides speaker diarization, turn detection, contextual and keyword biasing, and 25 evaluated languages with code-switching. It returns turn-level timestamps, not word-level timestamps.
• Segment Anything Model 3.1: An open-weight Meta model, also available on Model API, that returns boxes and pixel masks for objects named in a text prompt, and tracks objects across video frames.
• Open-weight models: Muse Spark Open is an 816-billion-parameter mixture-of-experts model with 42 billion active parameters, released as BF16 and NVFP4 checkpoints for self-hosting with SGLang or vLLM or through hosted providers. Muse Glimmer is a smaller multimodal model, distilled from Muse Spark, for local and edge use under an Apache 2.0 license. It runs on vLLM, SGLang, llama.cpp, or ExecuTorch.

Models and Pricing

Prices are Model API pay-as-you-go rates per 1 million tokens unless noted.

• Muse Wonder (muse-wonder-1.0): Built for complex software engineering and long-horizon agent work. Input $4.00, cached input $0.20, output $20.00. Muse Wonder Full, a higher-capability variant, is available only through Muse Code subscriptions.
• Muse Spark (muse-spark-1.3, muse-spark-1.2, muse-spark-1.1): Built for agentic and coding work at a lower cost than Muse Wonder. Input $1.25, cached input $0.15, output $4.25.
• Muse Spark Contributor variants (muse-spark-1.3-contributor, muse-spark-1.2-contributor): The same models with discounted prices. Input $0.10, cached input $0.002, output $0.20.
• Muse Ember (muse-ember-1.0): Built for high-volume, repeated work such as sub-agent tasks, routing, and classification. Requests up to 200K input tokens: input $0.10, output $0.50. Requests over 200K input tokens: input $0.30, output $1.50.
• Muse Image (muse-image-1.0): Image generation and editing, at $0.01 per generated image.
• Muse Voice Transcribe (muse-voice-transcribe-1.0): Speech-to-text, at $0.18 per hour of audio.
• SAM 3.1 (sam-3.1): Image and video segmentation, at $2.50 per 1,000 images and $0.20 per 1,000 video frames.
• Web search grounding: Adds $2.50 per 1,000 search queries.

Audience & Use Cases

• Audience: Software developers, AI and machine learning engineers, and platform teams building agents and applications.
• Use Cases: Multi-file coding changes in real repositories, multi-hour agent runs with sub-agents (the larger model plans and Muse Ember handles volume), classification and routing across thousands of records, live transcription for voice agents and call analysis, image generation, and object segmentation in images and video.

Technical Specifications

• API formats: Responses API (the default for agents, with reasoning carried across turns), Chat Completions (OpenAI-compatible), and Messages API (Anthropic-compatible). All three use the same models and bearer-token authentication (MODEL_API_KEY) and cost the same per token.
• Compatibility: Works with the OpenAI SDK, the Anthropic SDK, and OpenAI-compatible clients. Coding agents including OpenCode, Codex, and Claude Code can be pointed at Model API. Search grounding and tool search run only on the Responses and Messages APIs.
• Data handling: In Standard mode, customer content is not used to train Meta models. In Contributor mode, available on the Muse Spark 1.3 and 1.2 variants, content may be used to train future Meta models in exchange for lower prices.
• Availability: Muse Wonder is not available in South Korea.