Summary: Meta released Muse Spark Open on October 15, 2026, an 816-billion-parameter open-weights model aimed at coding and agent work, and showed it running on just two Mac Studios. The trick is a combination of MoE sparsity, 4-bit quantization, and Apple’s MLX framework with Thunderbolt 5 RDMA. For builders weighing local deployment, this is a real but expensive proof of concept — not a cheap data-center replacement.
What Meta shipped
Muse Spark Open uses a mixture-of-experts architecture: 816 billion total parameters, but only about 42 billion active per generation. It supports up to 1 million tokens of context and handles text, images, audio, and video. Weights are available for download, so teams can self-host and fine-tune.
On benchmarks Meta published, the model scored 61.8 on JobBench (agent work) versus NVIDIA Nemotron 3 Ultra’s 27.2, and 89.2 on MCP-Atlas, ahead of Reflection AI’s Beam (78.7) and GLM-5.3 (84.2). But it is not dominant across the board: on Terminal-Bench 2.1, its 76.4 trailed Beam (80.1), GLM-5.3 (83.9), and Kimi K3 (85). Keep in mind these numbers come from Meta’s own evaluations and haven’t been fully independently verified.
Why two Macs are enough
Three technical pieces make this work:
- Unified memory. A Mac Studio with the M5 Ultra chip can be configured with up to 512GB, and the integrated GPU accesses it directly — no per-card VRAM ceiling.
- Quantization. Per Apple Machine Learning Research, the BF16 weights need over 1.6TB of memory. NVFP4 four-bit quantization brings that down to roughly 493GB — still too big for one 512GB machine once you account for inference overhead, hence two machines.
- MLX with RDMA. Apple’s MLX framework uses Thunderbolt 5 remote direct memory access so each Mac holds part of the model and exchanges intermediate results, using tensor parallelism.
Apple measured about 24.99 tokens per second on two M5 Ultra Mac Studios at an 8,192-token input context, rising to 36.73 tokens with four machines. Meta also demoed code-editing tasks and video analysis in DaVinci Resolve on the setup.
This follows a pattern we’ve tracked with GLM-5.3’s weights — big MoE models are increasingly runnable on unusual local hardware, and GLM-5.3 running locally on 753B parameters showed the same trend from the Chinese side.
The deployment math isn’t simple
“Two Macs” sounds accessible; it isn’t yet. The M5 Ultra Mac Studio with 512GB of memory had not shipped as of October 15, and no price was announced. For reference, VentureBeat notes four Mac Studios with 256GB each total around $43,000. And the license has strings: under the Muse Model License, companies with over $20 million in annual revenue (including affiliates) that resell model outputs as MaaS must negotiate a separate agreement with Meta. Internal use and ordinary product features are unaffected.
The real value of local deployment here is control — owning the runtime, reducing API dependency, keeping data in-house. Whether it beats cloud inference on cost depends on your hardware outlay, throughput needs, workload shape, and maintenance capacity.
What I’d take from this
Don’t rush to budget for a Mac Studio cluster. But do note the direction: high-capability open weights plus consumer-adjacent hardware means deployment choices are no longer binary between a GPU server rack and an API meter. If your team runs open models today, the practical next step is benchmarking 4-bit quantization quality on your own workloads before assuming the full-precision numbers apply.