Meta open-weights Muse Spark Open, an 816B model built to run on two Mac Studios

15 October 2026. Meta releases the weights of Muse Spark Open, a multimodal model with 816 billion total parameters. 42 billion. The number of active parameters per token, thanks to a mixture-of-experts architecture that only activates a fraction of the model. 1 million. The context length, in tokens. Why it matters: Meta is repositioning its AI on the ground Chinese labs have dominated for two years — open-weight — and offering an unexpected deployment path: two networked Mac Studios instead of a data centre.

A « frontier » model you can actually download

Muse Spark Open is the open-weight variant of the Muse family, the frontier model Meta launched earlier this month. The difference comes down to one verb: download. Where closed models — Anthropic’s Claude, OpenAI’s ChatGPT — run behind an API, the weights of Muse Spark Open can be pulled from Hugging Face and modified locally.

Meta isn’t being modest. « It is the most capable US open-weight model to date, » wrote Alexandr Wang, the company’s head of AI, on X. The positioning is aimed squarely at Chinese rivals: Alibaba’s Qwen, Moonshot AI’s Kimi K3 and Z.ai’s GLM models have led the open-weight field for two years, and Meta wants the crown back with a model « built for coding and agentic work » that « excels at long-horizon tasks, complex software engineering, and multi-step reasoning ».

The architecture: 816B parameters, 42B active

The headline figure, 816 billion parameters, hides a subtler reality. Muse Spark Open is a mixture-of-experts: only 42 billion parameters are active per token, the rest sitting dormant in specialised experts the router calls on demand. That’s what makes the model usable on local hardware, despite a total size that would otherwise put it out of reach of a workstation.

The weights ship in BF16 and NVFP4, a 4-bit quantisation aimed at inference on recent silicon. On the tooling side, Meta provides the Helion kernels for inference and TorchTitan support for supervised fine-tuning and reinforcement learning. In short, this isn’t a bare weight dump — it’s a kit covering inference and adaptation, with extra safety post-training on top of the closed version.

The Apple bet: two Mac Studios, not a data centre

The real novelty is elsewhere. Meta says Muse Spark Open is optimised for Apple’s M5 Ultra chip, and that the model can run on two linked Mac Studios, forming a desktop « supercomputer ». It’s an unusual partnership between the two companies, and a strong signal for local inference.

The choice says something about where the market is heading. Two years ago, running a « frontier » model locally required a multi-GPU node. Today, Apple’s unified memory — hundreds of gigabytes reachable by the GPU — becomes a credible path to private inference: no data leaving the machine, no network latency, no API dependency. For sovereignty- or privacy-sensitive teams, that argument weighs heavily against all-in-cloud.

The license: open-weight is not open source

Read the terms before celebrating. Muse Spark Open ships under the Muse Model License, which requires attribution and, more importantly, imposes a separate agreement once a service built on the model exceeds $20 million in revenue. In other words, the weights are open for experimentation and self-hosting, but commercialisation at scale goes back through a negotiation with Meta.

That’s the distinction between open-weight and open source in the strict sense: you can read the weights, run them and adapt them, but the license doesn’t guarantee the unlimited commercial freedoms an OSI license would. The « open-weight » label is chosen carefully — Meta doesn’t claim open source, and that’s exactly what lets it keep a hand on downstream monetisation.

What to take away

On benchmarks the picture is mixed. Muse Spark Open leads on agentic tasks but trails rivals like Kimi K3 on some coding and reasoning exercises. Picking a model is therefore not about a single score — it depends on what you want to do, and where you want to run it.

For a homelab or a team that wants a frontier model locally, Muse Spark Open is the most credible US option right now — provided you have the hardware. For a commercial product headed past $20 million in revenue, the license becomes a cost to model from day one. And for pure coding, a recent Kimi K3 or Qwen is often more efficient per token.

Verdict

If you want frontier-grade local inference and have two M5 Ultra Mac Studios to hand, Muse Spark Open is the model to try first: it’s the best US open-weight model available, and the Apple silicon support is explicit. If you’re building a service you’ll sell, cost in the Muse Model License before you cross the $20 million threshold — the openness is real, the commercial freeness is not. If you only need to produce code without a locality constraint, stick with the best coding models of the moment; Muse Spark Open’s agentic edge doesn’t pay off everywhere yet.

References

• Meta AI Research — Introducing Muse Spark Open
• PyTorch — Muse Spark Is Now an Open-Weights Model
• Hugging Face — meta-models/Muse-Spark-Open-816B-A42B
• The Guardian — Meta releases Muse Spark Open, a new open-weight AI model (16 October 2026)
• Unite.AI — Meta’s 816B-Parameter Muse Spark Open Weights Are Now Public