Meta Releases Open-Weight Muse Spark Open - 816B MoE, Licensees Above $20M in Gross Revenue Need a Separate Agreement for Model-as-a-Service Use

Meta released the weights for Muse Spark Open on October 15, 2026. The 816B-parameter MoE (42B active) has a 1M-token context window. The custom Muse Model License is free for many uses but requires attribution, and a licensee whose 12-month gross revenue, including affiliates, exceeds $20M needs a separate agreement to use it in a Model-as-a-Service business. Self-hosting takes 700GB+ of GPU VRAM or several M5 Ultra Mac Studios.

Meta Muse open-weight license LLM Apple MLX

On October 15, 2026, Meta made the weights of Muse Spark Open public¹. It has 816B parameters in total with 42B active, handles multiple modalities natively, and accepts up to 1 million tokens of context. Meta calls it the most capable open-weight model from the US so far, and says it was built for coding and agent workloads.

Until now, Meta has offered Muse Spark through its own API and its coding agent. This release is the open-weight version, so the model can be placed on a company’s own servers or local machines. The license, however, is not a general-purpose one such as Apache 2.0. It is Meta’s own Muse Model License, and its terms depend on how the model is used and on the size of the business. The hardware needed to self-host is also substantial.

Where It Is Available

Seven providers offer API access from day one: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI¹. The model card also notes that API access comes through third-party inference providers². Besides Hugging Face, the weights can be downloaded from Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta lists SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth as supported inference engines.

On Hugging Face, users have to log in and review the license conditions before they can reach the weights².

According to the model card, the model takes text, images, video and audio as input and produces text, and its training data covered more than 100 languages. It uses a Mixture of Experts (MoE) design in which 8 of 256 experts are picked for each token. The reasoning effort setting has four levels (low, medium, high and xhigh), and the deepest level, xhigh, is the default. For hard problems, coding and agent work, the model card points to high or xhigh; when a quick, short answer is enough, it points to the two lower levels.

Three License Points to Check

In its blog post, Meta says the license carries “limited commercial terms” and that many developers can use the model at no cost¹. Reading the text itself, three provisions are the most likely to matter in practice².

1. Attribution. Any product or service using Muse Spark Open or a derivative must show “Built with Muse Spark Open” prominently, either in its user interface or in accompanying docs, the website or a settings menu.

2. What counts as a derivative. The license defines “Derivatives” to include models created, trained, modified (fine-tuning included) or distilled with Muse Spark Open, and also models built using its outputs or results. A separate model trained on Muse Spark Open outputs is therefore a derivative and falls under the attribution requirement and the usage policy.

3. A revenue test for Model as a Service. Section 3 covers businesses that give third parties access to model outputs through an API or similar interface (“Model as a Service”). If the licensee, together with affiliates, earns more than $20 million in gross revenue over any consecutive 12 months, they must notify Meta and sign a separate written agreement before using Muse Spark Open or a derivative in that business. The section does not cover internal use, research, other commercial activities outside Model as a Service, or using the model as a component inside the licensee’s own product (as long as that product is not itself Model as a Service).

The usage policy is also part of the license. It says the model is not meant for people under 18, prohibits uses such as military applications, weapons and running critical infrastructure, and bars deliberately circumventing safety measures. Meta also states that it does not claim ownership of outputs.

Tying conditions to revenue is not unique to Meta among large open-weight releases. The custom license for Qwen3.8’s 2.4-trillion-parameter model requires a separate license for Model as a Service or AI Work Assistant businesses whose combined revenue over 12 months exceeds $50 million. Muse Spark Open sets a lower bar of $20 million and applies it only to Model as a Service.

Hardware Needed to Self-Host

The model card puts the requirement for running it on your own system at 700GB or more of GPU VRAM for the NVFP4 version and 2TB for BF16².

Meta puts weight on running it on a Mac. The company claims that, within its class, no earlier model ran on a desktop setup of two M5 Ultra Mac Studios running the newest MLX release¹. Its explanation is that the M5 Ultra’s GPU can use as much as 512GB of unified memory per Mac Studio, so a pair of machines can hold the full model. Meta says it tried agent tasks entirely on-device with Muse Code running on Muse Spark Open, including changes to the iOS app of Ice Cubes (an open-source client for Mastodon) and operating DaVinci Resolve.

Apple published a technical post on distributed inference with MLX the same day³. By Apple’s figures, the Muse Spark Open weights take more than 1.6TB in BF16 and roughly 493GB even in 4-bit NVFP4. With the memory needed for computation during inference added, the model does not fit on one machine, so several Macs are linked with RDMA over Thunderbolt 5, a driver added in macOS 26.2. In Apple’s measurements, a cluster of four M5 Ultra Mac Studios produced more than 36 tokens per second with an 8,192-token context.

In other words, “runs on your desk” still means two M5 Ultra Mac Studios in Meta’s configuration, or more than 700GB of VRAM on GPUs. For many organizations, the first route to try it is likely to be one of the API providers listed above.

Meta’s Own Table Does Not Show a Single Winner

The model card includes a table comparing Muse Spark Open with other open-weight models of the same scale and architecture². All figures below are as published by Meta, with Muse Spark Open run at the xhigh reasoning effort.

Benchmark	Muse Spark Open	Qwen3.8-2.4T-A95B	GLM 5.3	Kimi K3
Artificial Analysis Intelligence Index	39	40	45	44
MCP-Atlas	89.2	84.5	84.2	82.3
Terminal-Bench 2.1	76.4	82.0	83.9	85.0
DeepSWE 1.1	60	57	67	68

Muse Spark Open has the highest MCP-Atlas score in the table, but on the Artificial Analysis Intelligence Index and on Terminal-Bench 2.1 and DeepSWE 1.1, which the model card files under Coding, GLM 5.3 and Kimi K3 score higher. Meta’s “most capable” claim is limited to open-weight models from the US, and that qualifier is worth keeping in mind when reading the table. Kimi K3’s full weights are also public, so a choice between open-weight models means comparing license terms as well as benchmark gaps.

After Muse Spark, Muse Code and Muse Wonder

In July, Meta launched its developer-facing Meta Model API alongside Muse Spark 1.1, and in August it released Muse Code, a terminal-based coding agent, in beta. On October 12 it announced Muse Wonder, a higher-end model offered through the Meta Model API and Muse Code. Muse Spark Open takes Muse Spark, which had been used through Meta’s services, and makes it available as weights that can be brought into a company’s own environment.

On safety, Meta says it assessed the most severe risks under its Meta Superintelligence Scaling Framework and that the model met the thresholds needed for release¹. Details are in the model card, and Meta says a full evaluation report is coming soon. The model card says the open-weight version went through extra safety post-training intended to make misuse harder, and recommends running it as part of a larger system with guardrails suited to the use case².

Before comparing performance, a company considering adoption needs to settle three questions: whether its use counts as Model as a Service under Section 3, whether any plan to train other models on the outputs falls under the derivative definition, and where it will meet the attribution requirement.