Meta Muse Spark Open: the 816B open-weight model, its Muse Model License and what it takes to run

On October 15, 2026, Meta Superintelligence Labs released Muse Spark Open, the first downloadable version of its Muse Spark model family, and the largest open-weight model in Meta's Muse line so far. It's a mixture-of-experts model with 816 billion total parameters, 42 billion of them active per token, and a 1-million-token context window. It ships under a new Muse Model License, which is closer to permissive than Llama's terms but isn't an open-source license. Meta calls it "the most capable US open-weight model to date," and says it can run locally on two Mac Studios with Apple's M5 Ultra chip.

This story is based on Meta's research and developer posts, the Hugging Face model card and license file, Artificial Analysis's model pages, Apple's Mac Studio announcement, and reporting from Reuters and TechCrunch, checked October 16, 2026.

The short version

• What: Muse Spark Open (meta-models/Muse-Spark-Open-816B-A42B), a sparse MoE model for coding and agent work. It takes text, images, video and audio, and outputs text.
• License: the Muse Model License. Free to use, modify, fine-tune, distill and sell, with a required "Built with Muse Spark Open" label and a revenue gate for model-hosting businesses.
• Hardware: four Blackwell Ultra GPUs, or two to four M5 Ultra Mac Studios with 1 TB of memory between them. Hosted APIs include OpenRouter, Together AI and Groq.
• Results: vendor-reported. Meta's own table puts it ahead of other US open models but behind the top Chinese ones on several tests.

What is Muse Spark Open?

Meta's Muse Spark models, launched in April, have been closed until now. Muse Spark 1.3 and the newer Muse Wonder models are still API-only. Until October 15, Meta's only open Muse model was the 30-billion-parameter Muse Glimmer, released under Apache 2.0 in August.

The Hugging Face model card describes a 62-layer MoE Transformer with 256 experts, eight active per token, plus a 1.8-billion-parameter encoder for images, video and audio. Reasoning effort runs from low to xhigh, the default.

Meta doesn't say which closed version it corresponds to. The Wall Street Journal reported in August that Meta planned to open Muse Spark 1.2.

What does the Muse Model License allow?

The license file on Hugging Face, dated October 15, 2026, grants "any person" permission to "use, copy, modify (including fine-tune), create derivative works, distill, merge, publish, distribute, sublicense, and/or sell" the model, free of charge. The conditions are:

• Branding. Any product or service using the model or a derivative must "prominently display 'Built with Muse Spark Open'" in its interface, documentation, website or settings menu.
• Derivatives include distilled models. A derivative is any model "created, trained, modified (including fine-tuned), distilled, or otherwise improved" using the model "or any outputs or results thereof," so models trained on its outputs carry the label too.
• The revenue gate. If you or your affiliates offer AI models to third parties via an API ("Model as a Service") and your combined gross revenue tops $20 million in any 12-month period, you must notify Meta and sign a separate agreement before offering Muse Spark Open that way. Internal use, research and embedding it in your own non-API product are exempt.
• A usage policy. Exhibit A bars unlawful uses, deception, passing off outputs as human-made, and circumventing safety measures, and says the model "is not intended for individuals under the age of 18."

How that compares with Meta's earlier terms, which our open-weight license comparison covers in detail:

Last verified: October 16, 2026 (license files on Hugging Face and Meta's sites)

Like Alibaba's and Mistral's revenue-gated licenses, it targets cloud resellers and API hosts, the businesses most likely to compete with Meta's own Model API. The gate applies to anyone offering any AI models as a service, so every large cloud counts. "Affiliates" isn't defined, and there's no termination or governing-law clause. The usage policy and branding rule also mean it isn't an open-source license under the Open Source Initiative's definition. This is a summary of the license text, not legal advice.

What hardware does it need?

Meta's developer post lists a BF16 checkpoint of about 1.6 TB and an NVFP4 build of about 530 GB that serves on "four Blackwell Ultra GPUs or eight H200s." AMD and Intel published their own quantized builds, and Unsloth's smallest GGUF needs about 300 GB.

The Mac setup is two M5 Ultra Mac Studios with 512 GB each, or four with 256 GB each, connected over Thunderbolt 5 with RDMA and running Apple's MLX framework. Apple's August 25 announcement said the 512 GB Mac Studio configuration was "coming in late October," with M5 Ultra models starting at $5,499. Meta's developer post says it "worked with RadixArk, Inferact, NVIDIA, Apple and Hugging Face" on launch support in inference software, and lists Apple among its runtime and silicon partners. TechCrunch's page summary says Meta and Apple "teamed up," but Apple hasn't commented and neither company has described a formal partnership.

How does it compare with other open models?

Meta's model card runs Muse Spark Open at xhigh effort and compares it with other open models at each model's own top settings. These numbers are vendor-reported, and Meta doesn't give settings for every rival:

Muse Spark Open leads on tool use (MCP-Atlas) and on the HLE-Diamond reasoning test without tools, and clearly beats Nvidia's Nemotron 3 Ultra. It trails the Chinese models on the composite index and on terminal coding, which is why the "US" in Meta's claim does real work. Reflection AI's Beam, another US contender, appears only on some tests. The Intelligence Index figures in this table come from Meta's model card. Artificial Analysis's own site now lists Muse Spark Open (xhigh) at 39 on its Intelligence Index v4.3.2, and its open-weights ranking shows GLM-5.3 (max) at 45 and Kimi K3 (max) at 44, the same as Meta's table (checked October 16). The index is a composite of Artificial Analysis's own evaluations, so it reflects that evaluator's test mix and settings. Artificial Analysis scored Meta's closed Muse Spark 1.3 at 48 on index v4.3 on October 12; that's an earlier index version, so the two numbers shouldn't be treated as one ranking.

Meta says the model got extra safety post-training and met its Superintelligence Scaling Framework release thresholds, but warns results may not hold for modified versions. Meta has published an evaluation methodology document; the full evaluation report is "available soon," its research post says.

What's confirmed, reported and uncertain

• Confirmed (Meta, Hugging Face, Apple): the October 15 release, the 816B/42B MoE design, 1M context, license terms, hardware configurations and partners; Apple's $5,499 starting price and late-October date for 512 GB Mac Studios.
• Third-party: the Intelligence Index score of 39, now listed on Artificial Analysis's site.
• Vendor-reported: every other benchmark and safety result above, and the "most capable US open-weight model" claim.
• Reported: Meta's August plan to open Muse Spark 1.2 (WSJ); a Meta–Apple team-up (TechCrunch metadata). Meta says only that it worked with Apple, among others, on launch software support.
• Uncertain: which Muse Spark version this is; how Meta will enforce the revenue gate; independent results beyond the Intelligence Index; when the full evaluation report ships.