Meta Releases Muse Spark Open: 816B MoE Model with Commercial License Requires Dual Mac Studios for Local Inference

Meta Releases Open-Weight Model Muse Spark Open: Free for Commercial Use, Requires Two 512GB Mac Studios for Local Execution

On October 15, Meta released Muse Spark Open. The model weights are available for direct download and can be used free of charge under the Muse Model License, including for commercial purposes. Meta claims it is "the strongest US open-weight model to date."

Since the launch of the Muse series, the flagship models Muse Spark and Muse Wonder (released three days prior) have only been accessible via Meta’s products and APIs. Only the smaller model, Muse Glimmer, had open weights. This marks the first time Meta has released weights for a large-scale Muse model.

The model has a total of 816 billion parameters but activates only 42 billion per token processed. This architecture is known as Mixture-of-Experts (MoE), which enables fast computation, though the entire model must still fit into memory. It supports a context window of 1 million tokens, accepts text, images, video, and audio inputs, outputs text, and focuses on coding and enabling AI to perform long-running autonomous tasks.

【Can You Download It? Can You Use It Commercially?】

Yes, you can download it. Weights are available on Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry. On Hugging Face, users must log in and complete an application form with their real name, date of birth, and full institutional name; submissions cannot be modified after submission.

Yes, commercial use is permitted. The license allows usage, modification, fine-tuning, distillation (training other models using its outputs), redistribution, and sale, all at no cost. There are three conditions.

First, products utilizing this model must prominently display "Built with Muse Spark Open" in the interface, documentation, website, or settings menu.

Second, users must adhere to the accompanying Acceptable Use Policy: the model cannot be used for military, weapons, critical infrastructure, or other high-risk applications; it cannot provide financial, legal, or medical professional services without proper licensure; and it must not target users under 18 years old.

Third, large companies selling model interfaces must sign a separate agreement. Specifically, if you sell access to this model or a derivative model via API or similar means, and your company (including affiliates) has generated over $20 million in total revenue in the past 12 months, you must notify Meta and sign a separate agreement. This threshold applies to total company revenue, not just revenue from this specific business line.

Embedding the model into your own app for features, internal corporate use, or research is exempt from the third condition, regardless of company size. While Llama’s previous license threshold was based on 700 million monthly active users, this new threshold targets revenue from model interface businesses.

【Can Standard Computers Run It?】

No. The model card specifies that the version compressed to NVIDIA’s 4-bit precision format (NVFP4) requires at least 700GB of VRAM, while the original BF16 version requires 2TB.

Meta states it has optimized the model for two M5 Ultra Mac Studio machines: each with up to 512GB of unified memory, combining to 1TB when linked, running via Apple’s MLX framework. Meta has not specified which compression version runs on Macs nor provided speed metrics.

Developers with server hardware can run it using mainstream inference frameworks such as SGLang, vLLM, TensorRT-LLM, and llama.cpp. Fine-tuning is also supported via PyTorch and TorchTitan.

Without such hardware, users can call APIs provided by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Pricing is determined by each platform; Meta has not published official rates.

【How Does It Compare to Other Models?】

Meta compared it only against open-weight models: Inkling from the US (produced by Thinking Machines, founded by former OpenAI CTO Mira Murati), NVIDIA Nemotron 3 Ultra, France’s Mistral Large 4, and China’s Qwen3.8, Zhipu GLM 5.3, and Moonshot Kimi K3. All following results are from Meta’s model card.

In Artificial Analysis’s Composite Intelligence Index, Muse Spark Open scores 39, compared to Inkling’s 26 and Nemotron 3 Ultra’s 23. The claim of being the "strongest US model" refers to this comparison. Among all open-weight models, it trails GLM 5.3 (45) and Kimi K3 (44), and is close to Qwen3.8 (40). Xiaomi’s MiMo-V2.6-Pro currently tops Artificial Analysis’s open-source leaderboard with 46 points but is not included in Meta’s comparison table.

It leads in tool calling and complex reasoning: scoring 89.2 on MCP-Atlas (which tests completing tasks via external tools through the MCP protocol), higher than competitors’ max of 84.5; and 26.1 on Human Last Exam (HLE, a set of extremely difficult interdisciplinary questions solved without tools), higher than competitors’ max of 22.2.

It is weaker in coding: scoring 76.4 on Terminal-Bench (command-line tasks), compared to Kimi K3’s 85.0 and GLM 5.3’s 83.9; and 60 on DeepSWE (software engineering tasks), compared to Kimi K3’s 68 and GLM 5.3’s 67.

Its most standout feature is jailbreak resistance. StrongREJECT measures whether the model outputs harmful content when induced; lower scores indicate better safety. Muse Spark Open scores 0.5, while other models score above 60. Meta stated that the open-weight version underwent additional safety training.

Compared to Meta’s own closed-source models, data cited by Meta three days ago from Artificial Analysis shows Muse Wonder scoring 52 and Muse Wonder Full scoring 56, placing Muse Spark Open significantly lower.

Original Article: https://t.co/SICKXtdZyV Official Blog: https://t.co/bMJ4GKl0An Model Card: https://t.co/oqcXeL9Ol0

↳ Quote @AIatMeta (AI at Meta): Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.

Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.

Key Model Specs: → 816B total / 42B active parameters → 1M token context window → Natively multimodal across text, image, video, audio