Meta Releases Muse Spark Open, 816B-Parameter Open-Weight Multimodal Model

By
Harris
-
36

Key Points

  • Meta released Muse Spark Open on Oct. 15, 2026, calling it the most capable U.S. open-weight model to date.
  • The natively multimodal mixture-of-experts model has 816 billion total parameters, about 42 billion active per token, and a 1-million-token context window.
  • Weights are on Hugging Face and major cloud model gardens; Meta says the model can run on two Apple M5 Ultra Mac Studios with the latest MLX framework.
  • Hosted APIs are listed through Together AI, Fireworks, Groq and other partners under the Muse Model License with limited commercial terms.

Meta on Thursday released the weights for Muse Spark Open, a large open-weight AI model the company describes as the most capable U.S. open-weight system to date. The Meta AI Research announcement, dated Oct. 15, 2026, presents Muse Spark Open as a natively multimodal mixture-of-experts model built for coding, agentic workflows and long-horizon tasks.

The release sits in the open-weight race that already shapes how U.S. developers pick models for local and cloud work. It is a software and weights drop, not a headset product: Meta’s separate Muse Charm VR glasses hardware coverage on LiveNewsWorld remains a different story.

What happened

According to Meta, Muse Spark Open has 816 billion total parameters, with about 42 billion active for each token, and a context window of up to 1 million tokens. Inputs include text, images, video and audio; the model outputs text. Meta says it was trained for coding and agentic work, including complex software engineering and multi-step reasoning.

Weights are available for download on Hugging Face and through Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, Meta wrote. Hosted inference is listed with Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

Meta also says Muse Spark Open is the first model in its class to run on a desktop setup of two Apple M5 Ultra Mac Studios using the latest MLX framework. The company cites Apple’s unified memory of up to 512 GB per Studio and high memory bandwidth as the reason two machines can hold the full model. Supported inference stacks named in the blog include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations called out for AMD, Apple, Intel and NVIDIA silicon.

Latest development

Hugging Face published a matching technical walk-through of the same release, describing Muse Spark Open as an open-weight frontier multimodal mixture-of-experts model and pointing developers to the Muse Spark Open collection and the Muse-Spark-Open-816B-A42B checkpoint. The HF write-up also notes BF16 and NVFP4 checkpoint formats and documents text, image, audio and video input paths, without publishing independent leaderboard scores that this story can cite.

Meta says the open-weight build ships with custom Helion GPU kernels in SGLang, citing 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs and about a 3 percent end-to-end speedup on AMD MI350. Those figures come from Meta; this article does not invent or restate unaudited third-party benchmark tables.

On licensing, Meta is releasing the weights under the Muse Model License with limited commercial terms, which it says allow many developers to use the model for free. Readers should check the license text for the commercial limits rather than treat the release as fully unrestricted.

Why it matters

Open weights change who can inspect, fine-tune and host a frontier-scale model. Meta’s claim that Muse Spark Open is the most capable U.S. open-weight model to date is the company’s own framing; independent labs and media will need time to publish comparable evaluations. For U.S. developers, the practical news is availability: downloadable checkpoints, cloud gardens and a named list of hosted API partners.

The Apple silicon angle also matters for teams that want strong local inference without a rack of GPUs. Meta’s two-Mac-Studio MLX claim puts a concrete desktop path next to the usual data-center recipes. That sits alongside broader AI infrastructure demand already visible in recent chip earnings coverage, including our look at TSMC’s latest quarterly results.

Separately, readers following Meta’s consumer hardware should not confuse this model drop with the company’s Muse Charm VR glasses story. Muse Spark Open is about model weights and inference; Muse Charm is a headset product line.

The release also lands in a week when free and open chatbot offers remain a live consumer topic, including our earlier coverage of GPT-6 Luna and free ChatGPT access claims. Those products are closed or partner-hosted stacks; Muse Spark Open is Meta’s open-weight answer aimed at developers who want to run or host the model themselves.

What happens next

Developers can pull weights from Hugging Face and the cloud gardens Meta listed, or call the hosted partners named in the research post. Meta points readers to PyTorch TorchTitan for further tuning and to MLX for Apple silicon, which it says can pull weights directly from Hugging Face.

On safety, Meta says Muse Spark Open received additional open-weight safety post-training and was evaluated under the Meta Superintelligence Scaling Framework. Detailed results, Meta wrote, are in the model card, with a fuller evaluation report said to be available soon. Until independent labs publish comparable numbers, treat Meta’s capability and safety claims as company statements.

Watch for third-party evaluations, license clarifications and any follow-on Muse-series checkpoints. This package does not include Meta Platforms stock moves; no live quote was attached.

Related live coverage

LiveNewsWorld does not have a dedicated AI Live Center. For live U.S. and world news channels, use our Watch Live news hub. Market and AI-trade context sits on our Markets Live page.

Sources