Meta Unveils Muse Spark Open for Local and Cloud Agentic Workflows
An 816B parameter mixture-of-experts model enables agentic engineering on Mac Studios and enterprise clusters.
1Frontier Open Weight Release
Meta released the open weights for Muse Spark Open, describing it as the most capable US open-weight model to date. The architecture uses a sparse mixture-of-experts design with 816 billion total parameters. It selectively activates 42 billion parameters per token via top-8 routing across 256 distinct experts.
2Native Multimodal Perception
A dedicated 1.8 billion parameter perception encoder processes interleaved text, images, video frames, and audio in a single pass. Eliminating segregated pipelines per modality lets the model diagnose software bugs directly from narrated screen recordings. It simultaneously inspects underlying codebases, logs, diagrams, and video walkthroughs within one unified developer session.
3Million Token Context Architecture
The model features a 1M-token context window designed to ingest entire software repositories without chunking tasks. It handles multi-file refactoring, legacy dependencies, and thousands of files simultaneously. Architectural mechanisms like goal conditioning and context compaction preserve execution trajectory and critical state across prolonged, multi-step agentic problem-solving sessions.
4Shift to Agentic Engineering
Unlike general coding models optimized for single-shot completions, Muse Spark Open targets complex engineering horizons. It plans, acts, observes, and self-corrects while coordinating subagents across desktop, mobile, and browser environments. The model generates explanatory commentary before each discrete action to provide developers full visibility into its operational reasoning.
5NVIDIA Blackwell Ultra Performance
Meta trained Muse Spark Open on NVIDIA GB300 NVL72 rack-scale systems integrating 72 Blackwell Ultra GPUs. Optimized NVFP4 checkpoints leverage native FP4 Tensor Cores, shrinking the 816B memory footprint. On this hardware, the system delivers up to 50,000 tokens per second per GPU and 600 tokens per user.
6Apple Silicon Partnership
In a rare cross-industry collaboration, Meta worked with Apple engineers over several months to optimize deployment on Apple silicon. Former Apple executive Daniel Gross reportedly played a major role in facilitating this partnership. Both Meta AI chief Alexandr Wang and CEO Mark Zuckerberg publicly highlighted the joint optimization efforts with Apple.
7Local Desktop Execution via MLX
Using Apple's open-source MLX array framework, the model runs locally across two M5 Ultra Mac Studios. Apple's unified memory architecture provides up to 512GB per machine, creating a combined 1TB workspace. MLX leverages RDMA over Thunderbolt 5, letting two interconnected desktop units operate effectively as a single unified supercomputer.
8Hardware Benchmarks and Costs
The 4-bit quantized version requires 493GB of memory to load. Apple benchmarks show M5 Ultra time-to-first-token outperforms M3 Ultra by 4.6x, while multi-studio clustering accelerates large-context processing by 1.8x. However, enterprise hardware costs remain steep, with standard 256GB Mac Studio configurations already exceeding $11,000 prior to dual-setup pairing.
9Real World Agentic Validation
Meta demonstrated practical on-device autonomy using its Muse Code tool on local hardware. The agent autonomously implemented bug fixes and built full releases for the open-source Mastodon client Ice Cubes. It also operated DaVinci Resolve end-to-end, reviewing footage frame-by-frame to identify strong moments and crop vertical video.
10Ecosystem and Licensing Terms
Weights are distributed on Hugging Face, SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry under the Muse Model License. The model ships with auto-tunable Helion kernels in SGLang, boosting GPU speeds by up to 3.95x. Hosted APIs are live on platforms including Together AI, Fireworks, Groq, and Baseten.