Summary: Meta released the weights for Muse Spark Open on October 15, 2026, calling it the most capable US open-weight model to date. It’s a 816B-parameter (42B active) multimodal model with a 1M-token context window, built for coding and long-horizon agentic work. For builders, the headline is practical: it runs on two Apple M5 Ultra Mac Studios, and hosted inference went live the same day.
What actually shipped
Three things landed at once, and each matters for a different kind of builder:
- Weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, under the Muse Model License with limited commercial terms. Meta says many developers can use it for free.
- Hosted inference on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. I covered the day-zero Together AI access and pricing separately if you’re evaluating the API route.
- Local runtimes: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations across AMD, Apple, Intel, and NVIDIA silicon.
The desktop angle is the real story
The detail worth your attention: two Mac Studios behave like one machine. The M5 Ultra’s integrated GPU can address up to 512 GB of unified memory per box, so two machines hold the full model, and the chip’s 1.2 TB/s memory bandwidth is fast enough to keep tokens moving. Meta paired this with an MLX adapter that pulls weights directly from Hugging Face.
This isn’t a lab demo, per Meta’s write-up. They ran agentic workflows entirely on-device using Muse Code, including shipping real iOS changes to Ice Cubes (an open-source Mastodon client) — bug fixes, features, full build — and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame and cropping the strongest moments vertically.
If you’ve been following the local-agent space — Muse Glimmer as a 30B model for local agents, for example — Spark Open is a different weight class. You’re no longer trading capability for the ability to run at home.
Performance notes for self-hosters
The open-weight release ships custom, auto-tunable Helion GPU kernels in SGLang. Meta reports kernel-level speedups of 1.03–3.95× on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350. Modest at the system level, but the kernels are in the box if you’re already serving large MoE models.
For customization, Meta points to PyTorch’s TorchTitan for further tuning. Benchmarks are covered in their evaluation methodology report; I’m not restating score tables here without reading the methodology myself.
Safety and licensing, briefly
Meta says the model received additional safety post-training and was evaluated under its Superintelligence Scaling Framework, meeting release thresholds. Detailed results are in the model card, with a full report promised later. The license is the Muse Model License with limited commercial terms — worth reading closely before you build a product on it, since “free for many developers” is not the same as Apache-style licensing.
The takeaway
The interesting constraint just moved. Frontier-class agentic capability no longer requires a data-center rack to run privately — a pair of Mac Studios will do, and the API fallback exists everywhere. The open questions are the license terms for your specific commercial use, and whether real-world on-device throughput holds up outside Meta’s own demos. If you’re curious, the fastest next step is pulling the weights through MLX and running one of your own agentic workflows against them.