Days after launching its closed frontier models, Meta has returned to open weights. On October 15, 2026, the company released Muse Spark Open, an 816-billion-parameter mixture-of-experts model that Meta calls the most capable US open-weight model to date. Built for coding and agentic work and optimised to run on a pair of Apple Mac Studios, the release signals that Meta intends to compete at both ends of the market: closed models at the frontier, open models for developers.
Model Architecture and Specifications
Muse Spark Open is a mixture-of-experts (MoE) model with 816 billion total parameters, of which 42 billion are active for any given token. According to Meta's research blog and its Hugging Face model card, the key specifications are:
- Context window: 1 million tokens
- Modalities: natively multimodal input across text, images, video and audio, with text output
- Structure: 62 layers with a hidden dimension of 8,192, and 256 experts holding roughly 97% of all parameters
- Lineage: it shares a base model with Muse Spark 1.3, the closed lightweight model in Meta's lineup
Meta says the open release received additional safety post-training to make it more robust against misuse when deployed as a standalone model without the system-level guardrails used in Meta's own products. The company evaluated it under its Meta Superintelligence Scaling Framework and says it met the framework's release thresholds, with a full evaluation report promised soon.
Built to Run on a Mac
The most unusual aspect of the launch is its partnership with Apple. Meta says Muse Spark Open runs locally on two M5 Ultra Mac Studios using Apple's MLX framework. Each machine offers up to 512GB of unified memory, and Meta says the pair behaves like a single system with roughly a terabyte of memory.
That still puts local use out of reach for most individuals. Engadget estimates that an alternative configuration of four 256GB M5 Ultra Mac Studios would cost roughly $38,000, and Apple has not yet announced pricing for the 512GB model, which is slated for late October. On servers, the 4-bit quantised version requires about 700GB of GPU memory.
To demonstrate on-device agentic workflows, Meta paired the model with Muse Code, its AI coding terminal. Examples included bug fixes, new features and a full build for an open-source Mastodon client on iOS, as well as frame-by-frame footage review and cropping in DaVinci Resolve.
Benchmarks and Caveats
Meta's own evaluations claim Muse Spark Open outperforms two leading Chinese open-weight models, GLM 5.3 and Kimi K3, on agentic tasks and visual reasoning. TechCrunch reported that Meta's results show the model ahead of other popular open-weight models, while Technology.org and Engadget both noted that these benchmark figures have not been independently verified.
Meta's announcement also highlights infrastructure work: custom Helion GPU kernels in SGLang deliver kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD hardware, with a 3% end-to-end gain on AMD's MI350.
Availability and Licensing
Muse Spark Open is broadly distributed from day one:
- Weights: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models
- Hosted APIs: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI
- Inference engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA silicon
- Fine-tuning: supported through PyTorch and TorchTitan, including reinforcement learning
The model ships under the Muse Model License, which carries limited commercial terms. Meta says those terms allow many developers to use the model for free, though it is not a standard permissive open-source licence, and organisations should review the conditions before commercial deployment.
Why It Matters
Muse Spark Open arrives just three days after Meta released Muse Wonder and Muse Wonder Full, its first frontier-class models, neither of which offers downloadable weights. That decision raised questions about whether Meta was abandoning the open strategy that made Llama so influential. This release suggests a two-track approach instead: keep the most capable systems closed and monetised, while offering a strong open model to retain developer mindshare.
It also sharpens the competition with Chinese labs, which have dominated the open-weight leaderboard for much of the past year with models such as Kimi K3 and GLM 5.3. A credible US-built alternative at this scale gives enterprises with data-sovereignty or supply-chain concerns another option they can self-host.
Finally, the Apple partnership points to a future in which frontier-scale models run on high-end desktop hardware rather than only in data centres. For now, a terabyte of unified memory remains an enterprise-level investment, but it is a notable step toward truly local, private AI for demanding coding and agent workloads.
What to Watch
The key tests ahead are independent benchmark results, the promised full safety evaluation report, and how quickly developers adopt the model on hosted platforms versus local hardware. Pricing and availability of Apple's 512GB Mac Studio will also determine how practical the on-device pitch really is.
