Meta Releases Muse Spark Open with 816B Parameters
Meta released Muse Spark Open, an 816B open MoE model with native multimodality and a 1M token context window built for agentic coding.
The open MoE model pairs 1M context tokens with native perception for autonomous software engineering workflows.
Meta’s artificial intelligence division released the files and weights for Muse Spark Open on Thursday (October 15, 2026). Engineered specifically for software engineering and complex technical reasoning, the system packs 816 billion total parameters into a Mixture-of-Experts (MoE) design. It represents the company’s largest open-weights release to date, built to power autonomous agents tackling massive codebases.
Inside the Muse Spark Open Architecture
Deploying an 816-billion-parameter system requires strict control over computational overhead. Meta organized the weights across 62 layers with a hidden dimension of 8192, yet the model activates only 42 billion parameters for each generated token. A routing mechanism directs queries to 8 active experts out of 256 on every forward pass. This setup preserves low latency relative to dense models of comparable scale, while maintaining a context window of 1 million tokens.
Perception is integrated directly into the core rather than grafted on through an external computer vision network. Spark Open incorporates a native 1.8-billion-parameter perception encoder capable of handling interleaved streams of text, still images, audio clips, and video recordings. In daily engineering workflows, this allows an agent to simultaneously inspect user interface screenshots, system architecture diagrams, git diffs, and source files during a single debugging run.
Cloud Distribution, Formats, and Enterprise Clusters
The weights landed on Hugging Face in two formats: uncompressed BF16 (around 1.5 TB in safetensors) and NVFP4, a native 4-bit quantization tuned for NVIDIA’s Blackwell silicon. According to NVIDIA’s engineering blog, enterprise setups running GB300 NVL72 racks and standalone B200 GPUs can deploy the model through TensorRT-LLM, vLLM, and SGLang. Open-source inference tools such as llama.cpp and Unsloth also received day-one support.
Alongside managed clusters on Google Cloud Model Garden, Microsoft Foundry, and Amazon SageMaker AI, Meta lined up hosted API options. Services including Groq, Together AI, Fireworks AI, Baseten, Nebius, Modal, and OpenRouter made endpoints available at launch.
Local Mac Studio Deployment and License Terms
For engineering organizations barred from piping proprietary source code to external servers, the documentation outlines an on-premises deployment path. Apple’s MLX library was updated to distribute inference across two interconnected Mac Studio workstations fitted with M5 Ultra processors. The technical feasibility and memory loading steps are confirmed. However, real-world throughput figures under sustained production workloads have not yet been verified by independent developers and remain categorized as a RUMOR.
Distribution follows the Muse Model License. The terms match Meta’s previous open-release strategy: free access for academic research and standard software engineering, paired with contractual restrictions that limit direct commercial competitors with massive active user bases. Meta describes Spark Open as the most capable open-weights reasoning and coding model developed in the United States, though third-party community evaluations on benchmarks like SWE-bench Verified remain ongoing.
Sources
- X (@AIatMeta) - Official announcement of Muse Spark Open
- Meta Developer Platform - Technical overview and Muse Model License terms
- Hugging Face Blog - Release documentation and MoE weights repository
- NVIDIA Technical Blog - Inference analysis on Blackwell GB300 NVL72 and perception encoder details
- 9to5Mac - Report on local execution for Mac Studio systems with M5 Ultra