What are the architectural specifications of Muse Spark Open?

It uses a sparse mixture-of-experts architecture with 816 billion total parameters, 42 billion active parameters per token, 256 experts per MoE layer, and 8 active experts per token.

Meta Muse Spark Open API Pricing & Specs

Q: What is the pricing for the Meta Muse Spark Open API?

The API is priced at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens.

Muse Spark Open API Price and Benchmarks. A macro shot of a gold-wire bonded silicon die resting on a dark slate surface, illuminated by a sharp amber sidelight reflecting off microscopic circuit traces — GPUniq

Data as of October 16, 20265 sources

Drafted with AI tools and edited by a human. The figures and conclusions were checked by a GPUniq editor.

Meta prices the Muse Spark Open API at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens.

Token Type	Price, $/1M tokens
Input tokens	1.25
Output tokens	4.25

Hosted API providers include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

Architecture Specs

Muse Spark Open uses a sparse mixture-of-experts architecture with 816 billion total parameters and 42 billion active parameters per token. Each MoE layer contains 256 experts, using 8 active experts per token.

Self-hosting the full BF16 checkpoint demands 2 TB of VRAM. You can use the Muse-Spark-Open-816B-A42B-NVFP4 format to reduce this footprint, though you still must load the full 816B parameters into memory. Commercial deployment is subject to the Muse Model License.

Context Window Limits

The model supports a maximum context window of 1,048,576 tokens. Meta built this for multi-file codebases, long documents, and multi-step reasoning.

Meta does not specify a maximum output token length, meaning extremely long outputs might experience constraints.

Benchmark Scores

LuminaBench reports an 89.2 on MCP-Atlas for agentic, knowledge, and reasoning tasks. On Terminal-Bench 2.1 for coding evaluations, it reaches 76.4.

Benchmark	Score	Category
MCP-Atlas	89.2	Agentic and Reasoning
Terminal-Bench 2.1	76.4	Coding
SimpleVQA with tools	73.6	Multimodal
JobBench	61.8	Tool Use

Deployment Options

You can download weights from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Runtime frameworks include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth.

Hardware support covers NVIDIA, AMD, Intel, and Apple systems, including Mac Studio units with M5 Ultra chips.

Methodology

Facts are compiled from the vendor's own announcement and documentation, linked below, at the time of writing; where the vendor hasn't published a number it is marked as such. GPUniq prices, when the model is already available here, come from our live catalog.