Meta Releases Muse Spark Open: 816B-Parameter Multimodal Model for Local Mac Clusters

The Three-Question Framework

What Happened?

On October 15, 2026, Meta Superintelligence Labs released open weights for Muse Spark Open, an 816-billion-parameter natively multimodal mixture-of-experts model activating 42 billion parameters per token with a 1-million-token context window.

Why It Matters

Muse Spark Open represents the largest US-developed open-weight model to date, demonstrating that high-capacity agentic and software engineering reasoning can run locally on two Apple M5 Ultra Mac Studios pooling 1 terabyte of unified memory.

What Could Happen Next

Developer communities are implementing downstream quantization adapters across llama.cpp and vLLM, while enterprise cloud hosts roll out dedicated serving endpoints under the commercial terms of the Muse Model License.

Executive Key Takeaways

Muse Spark Open pairs 816B total parameters with 42B active parameters per token across a native 1,048,576-token context window.

Runs on desktop hardware using two connected Apple M5 Ultra Mac Studios with 512GB RAM each via Apple MLX, tapping 1.2 TB/s unified memory bandwidth.

Includes custom Helion GPU kernels integrated into SGLang, delivering 1.03x to 3.95x kernel speedups on NVIDIA and AMD hardware and 3% end-to-end acceleration on AMD MI350.

Now accessible via API providers including OpenRouter (meta/muse-spark-open at $1.25 per million input and $4.25 per million output tokens).

Source: In an official technical disclosure published on October 15, 2026, Meta Superintelligence Labs introduced Muse Spark Open, releasing complete model checkpoints on Hugging Face alongside hosted API endpoints.

The model architecture is built from the ground up for agentic execution and multi-step tool invocation across text, audio, image, and video modalities. In internal validation tests conducted by Meta, the model completed autonomous software engineering tasks on Ice Cubes, an open-source Mastodon client, implementing full features, bug fixes, and clean builds without human code modifications.

For creative automation, researchers verified end-to-end video timeline orchestration in DaVinci Resolve, directing the model to inspect video footage frame by frame to identify key highlights and automatically crop sequences into vertical aspect ratios.

To enable local deployment without server racks, Meta partnered with Apple's MLX engineering team. Because each M5 Ultra workstation provides up to 512 gigabytes of unified memory, clustering two desktop systems over a high-bandwidth interconnect allows the entire uncompressed model weight distribution to fit inside a single shared memory pool.

What remains to be determined is whether open-source quantization researchers can compress the 816B parameter structure down to single-machine 192GB unified memory configurations without severely penalizing agentic tool-use accuracy.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations