Meta Releases Muse Spark Open, 816B Parameters Under a Licence It Calls Limited

Meta released the weights of Muse Spark Open on Oct. 15, a model with 816 billion total parameters, 42 billion active, and a 1 million-token context window, according to Meta's announcement. The licence is the Muse Model License, which Meta describes as carrying "limited commercial terms."

Meta calls the model "the most capable US open-weight model to date." That is Meta's own claim. The announcement has a key benchmarks section and links an evaluation methodology report, but the page text carries no scores and does not say who ran the evaluations. Engadget reported that Meta's own benchmarks show the model ahead of GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning, again without figures.

What it takes to run Muse Spark Open

Meta said the model runs on two Apple M5 Ultra Mac Studios using MLX. Each machine has up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth, and Meta says the pair "behave like one supercomputer with a terabyte of unified memory."

Engadget reported that four 256 GB machines also work, at roughly $38,000 for the setup. Apple has not announced pricing or a ship date for the 512 GB Mac Studio, which Engadget says is expected in late October.

Spec	Figure	Source
Total parameters	816 billion	Meta
Active parameters	42 billion	Meta
Context window	1 million tokens	Meta
Local setup	2 x 512 GB M5 Ultra Mac Studio	Meta
Alternative setup	4 x 256 GB, about $38,000	Engadget

Engadget describes the model as 800 billion parameters. Meta's own materials say 816 billion.

Where the weights are

Meta lists downloads on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted APIs come from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Supported inference engines include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth.

Hugging Face's public model index lists meta-models/Muse-Spark-Open-816B-A42B with a licence tag of "other", 46 likes and 402 downloads at the time of retrieval. An NVFP4 variant under the same namespace shows 331 downloads. The index shows the repository was created on Sept. 15, a month before the announcement, and does not show when it became public. It does not show whether access is gated.

Meta also said custom Helion GPU kernels in SGLang gave 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350. Those are Meta's figures.

What the post leaves out

The announcement makes no statement about training data. It says the licence lets many developers use the model for free, and the text retrieved does not set out where the commercial limits sit. Liquid AI's LFM2.5-24B-A2B shows what a published threshold looks like: its licence requires a commercial agreement from companies with $10 million or more in annual revenue. Meta's post gives no equivalent figure for the Muse Model License. Google's EmbeddingGemma 2 is covered on the same desk.

Meta said it added safety post-training and evaluated the model under its Meta Superintelligence Scaling Framework, and that it met the framework's release thresholds. It said a full evaluation report is "available soon" and gave no date.