Apple benchmarks Muse Spark Open on Mac Studio with MLX and RDMA
Apple’s Machine Learning Research team demonstrated how its MLX framework enables running large language models locally on Mac Studio hardware. The team benchmarked three frontier models—Muse Spark Open (816B parameters), Mistral Medium 3.5 (128B), and Qwen 3.8 (27B)—on both M3 Ultra and M5 Ultra Mac Studio configurations. Muse Spark Open, a mixture-of-experts model, requires distributed…
Key points
- MLX on Mac Studio runs Muse Spark Open (816B params) via distributed inference across multiple Macs
- M5 Ultra’s GPU Neural Accelerators cut time-to-first-token by up to 4.6× vs. M3 Ultra
- MLX supports RDMA over Thunderbolt 5 for low-latency multi-Mac communication
The research highlights that Qwen 3.8 and Mistral Medium 3.5 fit on a single Mac Studio with 512GB of RAM, while Muse Spark Open needs parallelization. The team also showed how MLX’s mlx.launch and mlxlm.server tools enable distributed inference and agentic AI workflows locally, with examples using OpenCode. The benchmarks emphasize Apple silicon’s efficiency for LLM inference, with power consumption as low as 200W for M3 Ultra and 300W for M5 Ultra.
Model page: Muse Spark Open →
The story so far
2 episodes →- Apple benchmarks Muse Spark Open on Mac Studio with MLX and RDMAthis story
Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
Apple Machine Learning Research · 15 October 2026
Loading the full article…
This text was published by Apple Machine Learning Research. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
7sources- Hacker News discussion · 31 pointsnews.ycombinator.com
- Meta’s 816B-Parameter Muse Spark Open Weights Are Now PublicPress · Unite.AI ·
- Meta says it worked with Apple to make Muse Spark Open run locally on M5 Ultra Mac Studios, letting users pair two to four Studios to code or do agentic workPress · cnbc.com ·
- Meta releases Muse Spark Open, saying it is the most capable US open-weight model, with 816B total and 42B active parameters and a 1M-token context windowPress · research.meta.ai ·
- Meta’s Muse Spark Open AI model can run locally on Apple Mac StudioPress · cryptobriefing.com ·
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from MetaPrimary source · Hugging Face ·
- Meta built a huge open-weight AI model for Macs, but you'll need 1TB of RAM to run itPress · engadget.com ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Hardware & Compute
All →- Nvidia CEO claims only the company can build $100B AI factories · 14 src
- CoreWeave seeks insurers’ capital to fund chip spending · 2 src
- NVIDIA launches RTX Spark PC for creators and AI agents · 1 src
- User reverse-engineers Plustek scanner driver with Claude’s AI assistance · 1 src
- Hugging Face launches Llama Windows for local AI on Windows 11 · 1 src
Comments
via GitHub Discussions