At 2026-10-16 Beijing time, Meta publicly released the model weights for Muse Spark Open, while Apple on the same day issued a technical note detailing how it runs the model using MLX across multiple Mac Studios. Muse Spark Open contains 816 billion parameters, with 42 billion activated per inference and a context window of one million tokens. Meta refers to it as the most powerful open-weight model in the United States to date, though this is a company self-assessment; no independent or unified evaluation has confirmed this ranking. Apple’s report also clearly delineates hardware boundaries: the model cannot be loaded into a single Mac, and demonstration configurations require at least two M5 Ultra Mac Studios.
Open Weights, Licensing, and Model Specifications
Open weights do not equate to open source. Meta released the weights under the Muse Model License, explicitly stating that the license includes limited commercial terms, thus developers receive capabilities for downloading, running, and customizing the model, but not unrestricted commercial use authorization. The weights are now available via Hugging Face and included in Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Meta also listed hosting inference platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI; local deployment and cloud invocation represent two distinct paths.
According to Meta, Muse Spark Open is a native multimodal mixture-of-experts model designed primarily for programming, agents, long-range tasks, complex software engineering, and multi-step reasoning. 816 billion refers to the total parameter count, while 42 billion denotes the number of parameters activated during processing of a single input—these should not be conflated as a single computational load. One million tokens is the maximum context specification provided by Meta, which cannot be directly applied to Apple’s entire speed profile. Meta also states that the model can run on inference engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth, with optimizations tailored for AMD, Apple, Intel, and NVIDIA hardware.
Why One Mac Still Can’t Fit It
The most concrete new information in this release comes from Apple’s memory calculations. Apple states that when Muse Spark Open uses BF16 native precision, the model parameters alone require over 1.6TB of memory; switching to NVFP4 quantization reduces the model size to approximately 493GB. Inference also requires storing intermediate computation data, so even a single M5 Ultra Mac Studio with 512GB unified memory cannot independently accommodate the full workload. Meta’s hardware documentation lists Apple configurations as either two Mac Studios each with 512GB unified memory or four Mac Studios each with 256GB unified memory—both of which constitute multi-machine workstation clusters.
Apple outlines two cross-machine splitting methods. Pipeline parallelism assigns earlier and later layers of the model to different Macs, where the first machine completes its layer computations before passing intermediate results to the next; tensor parallelism distributes the width of each layer across multiple machines, enabling concurrent execution of the same layer’s computation. The latter generates a large number of small communication operations, thus relying more heavily on low-latency links. Apple's tests used Thunderbolt 5 RDMA drivers introduced in macOS 26.2 and jaccl communication collections provided by MLX version 0.30.1 or higher; each pair of Macs is connected via two Thunderbolt cables to increase link bandwidth.
Official Benchmarks and Local Service Boundaries
Apple’s speed figures come with clearly defined test conditions: an input context of 8192 tokens followed by generation of 128 tokens. Under these conditions, the two-machine pipeline configuration had a first-token time of 8.4 seconds and a decoding speed of 15.74 tokens per second; the two-machine tensor parallelism achieved 8.2 seconds and 24.99 tokens per second; and the four-machine tensor parallelism reached 4.8 seconds and 36.73 tokens per second. These results were obtained through official tests using mlx_lm.benchmark, demonstrating performance under listed hardware, quantization methods, context, and parallel configurations but not extrapolated to a one-million-token context or all agent tasks.
Apple also showcased a local service interface: developers can launch mlx_lm.server across multiple Macs using mlx.launch, then connect tools like OpenCode to a local endpoint compatible with the OpenAI protocol. This example illustrates that model inference services can be deployed on operator-owned devices, but it does not imply that every tool, data source, or workflow associated with the model will automatically function offline. Meta’s release also retains multiple hosted API options; developers must still choose where to run based on data control, hardware investment, operational complexity, and application needs. Neither company disclosed user adoption rates, complete deployment costs, or cost-saving figures.
Security Disclosure and Next Steps for Observation
Conclusions regarding security and performance also need to be kept within source boundaries. Meta states that Muse Spark Open underwent additional safety fine-tuning and was evaluated according to the Meta Superintelligence Scaling Framework, achieving the release threshold required by the framework; detailed results are included in the model card, while a full evaluation report remains marked for later publication. Meta also notes that its self-tuned Helion GPU kernel achieved 1.03 to 3.95 times kernel-level acceleration on NVIDIA and AMD hardware, with a 3% end-to-end acceleration on AMD MI350. These are data disclosed by Meta, not Apple’s Mac benchmarks, and have not been independently audited in existing materials.
Editorial Analysis: This joint timing shows that the competition for deploying open-weight large models has evolved from simply whether files can be downloaded to quantization formats, unified memory, cross-machine communication, and local runtime coordination. It offers new self-hosted paths for research institutions and teams with professional equipment but does not make an 816-billion-parameter model easily runnable on ordinary personal computers. Next steps should observe independent capability and security evaluations, the actual constraints of limited commercial terms in the Muse Model License, total costs of multi-machine configurations, and whether quantization and runtime optimizations can continue to lower hardware thresholds.