How much does Muse Spark Open cost to use?

The weights are free to download under the Muse Model License. Hosted APIs list $1.25 per million input tokens, $0.15 cached and $4.25 per million output tokens on Together AI and OpenRouter as of October 16, 2026, the same rate as Meta's hosted Muse Spark 1.3. Together AI offers it free until October 18, 2026 at 12 p.m. PT. It is also served by Baseten, Fireworks, Groq, Modal and Nebius.

Is Muse Spark Open free for commercial use?

Mostly. The Muse Model License allows use, modification, fine-tuning, distillation and redistribution, including commercially, but requires a visible 'Built with Muse Spark Open' notice and compliance with Meta's usage policy. Companies offering the model as an API service with more than $20 million in revenue over 12 months must sign a separate agreement with Meta. It is not an OSI open-source licence.

What hardware do you need to run Muse Spark Open?

Meta's model card says at least 700 GB of GPU memory for the NVFP4 weights and 2 TB for BF16, with SGLang, vLLM, TensorRT-LLM, Transformers, Unsloth or llama.cpp. Meta also optimised it to run on two Apple Mac Studios with M5 Ultra and 512 GB of unified memory each, using MLX.

Quick Answer

What Is Muse Spark Open? Meta's 816B Open-Weight Model

Q: What is Muse Spark Open?

Muse Spark Open is Meta's open-weight AI model, released on October 15, 2026. It is a natively multimodal mixture-of-experts model with 816 billion total parameters (42 billion active per token), a 1-million-token context window and text, image, video and audio input. Meta built it for coding and long-horizon agentic work and calls it the most capable US open-weight model to date. Weights are on Hugging Face in BF16 and NVFP4.

Published: October 16, 2026

The short answer

Muse Spark Open is Meta’s first open-weight model since it moved from Llama to the Muse family: 816B parameters, 42B active, a 1M-token context and native multimodal input, released October 15, 2026. Meta built it for coding agents and long tasks. It is the strongest US open-weight model, but on Artificial Analysis’s Intelligence Index (39 per Meta’s own model card) it still trails Chinese open models such as GLM-5.3 (45) and Kimi K3 (44). The licence is permissive for most companies, with a revenue cap for model-as-a-service businesses.

Key facts

	Muse Spark Open
Released	October 15, 2026 (Meta Superintelligence Labs)
Architecture	Mixture-of-experts, 256 experts, 8 active per token; 62 layers
Parameters	816B total / 42B active
Context	1M tokens
Input / output	Text, image, video, audio → text
Languages	Trained on data from 100+ languages
Weights	Hugging Face (BF16, NVFP4), SageMaker AI, Google Cloud Model Garden, Microsoft Foundry
Hosted APIs	Together AI, OpenRouter, Fireworks, Groq, Baseten, Modal, Nebius
API price	$1.25 input / $0.15 cached / $4.25 output per 1M tokens (Together, OpenRouter)
Licence	Muse Model License (attribution; $20M revenue cap for API resellers)

How it compares (Meta’s model card, xhigh effort)

Benchmark	Muse Spark Open	GLM-5.3 (max)	Kimi K3 (max)	Qwen3.8 Max	Mistral Large 4
AA Intelligence Index	39	45	44	40	38
JobBench (agentic)	61.8	61.4	54.3	52.7	—
MCP-Atlas	89.2	84.2	82.3	84.5	—
OSWorld Verified	80.4	—	84.8	—	—
DeepSWE 1.1	60	67	68	57	62
Terminal-Bench 2.1	76.4	83.9	85.0	82.0	—
HLE-Diamond (no tools)	26.1	16.4	22.2	17.3	—

Read the table as: best-in-class on tool use (MCP-Atlas) and job-style agent tasks, behind GLM-5.3 and Kimi K3 on pure coding benchmarks. These are vendor-reported scores; wait for independent runs before betting a product on the gaps.

The licence in plain terms

Allowed: use, copy, fine-tune, distil, merge, redistribute and sell products built on it.
Required: display “Built with Muse Spark Open” in the product, its docs or settings; follow the usage policy (no under-18 audiences, no malware, no unlicensed professional advice and similar).
Revenue cap: if you sell access to the model or a derivative as an API (“model as a service”) and your group’s revenue exceeds $20 million in 12 months, you need a separate agreement with Meta. Internal use and powering your own product features are exempt.
Hugging Face access is gated: you submit your full legal name and organisation.

Running it yourself

You need roughly 700 GB of GPU memory (NVFP4) or 2 TB (BF16) — an 8-GPU server class machine. Meta’s headline option is two Mac Studio M5 Ultra machines with 512 GB each linked through MLX; the M5 Ultra Mac Studio starts at $5,499 with 96 GB, and the 512 GB configurations cost far more and were still shipping slowly in October 2026. For most teams, the hosted APIs at $1.25 / $4.25 are cheaper until volume is very high.

Should you use it?

Yes if you need a US-origin open model for compliance, want to fine-tune or distil a strong agentic model, or run tool-heavy agents.
Compare first if raw coding score per dollar matters: GLM-5.3 Flash and DeepSeek V4.1 Flash cost far less per token — see current API prices and the best open-weight models ranked by value.

For Meta’s closed Muse Spark line, see Muse Spark capabilities and access.

Last verified: October 16, 2026.