Meta releases 816B Muse Spark Open for local Mac Studio deployment

Meta AI Research released Muse Spark Open, an open-weight Mixture-of-Experts model with 816 billion total parameters. The system activates 42 billion parameters per token. It accepts text, image, video, and audio inputs. Meta states it is the most capable US open-weight model currently available. It supports a one-million-token context window.

Root cause

The release provides an open-weight alternative to proprietary models. It targets developers requiring local inference capabilities. The architecture uses 256 experts with top-k routing activating eight per token. It includes a separate 1.8 billion-parameter Perception Encoder for visual input. The vocabulary contains 202,048 tokens. Training data covers more than 100 languages.

Attack path

This is not a security vulnerability or attack. It is a software release. The model received extra safety post-training. Meta states this improves robustness without system-level guardrails. It is designed for coding and long-horizon agentic work. It does not describe a threat vector or exploit.

Affected versions

The model is available on Hugging Face in BF16 and NVFP4 formats. It runs on two Apple M5 Ultra Mac Studios with 512 GB each via MLX. It also supports four Blackwell Ultra GPUs or eight H200s with NVFP4. Smaller setups of about 300 GB can use Unsloth 2-bit GGUF.

Mitigation

• Verify hardware compatibility before deployment.
• Monitor resource usage during inference.
• Review safety post-training documentation.
• Test agentic tool use in isolated environments.

What it means for organisations

Teams should assess local infrastructure requirements for large models. Check whether existing hardware supports the memory needs. Review data privacy policies for local processing. Ensure security teams understand the model’s safety features. Coordinate with development leads on integration plans. Evaluate vendor lock-in risks with open weights.

What to do and how to stay safe: Meta

• Audit current server capacity for large model loads.
• Test the NVFP4 build on available hardware.
• Review agentic tool permissions in development environments.
• Plan for ongoing monitoring of model performance.

Step-by-step guide: Prevent Prompt Injection Attacks: Practical Defences for AI Systems

General security guidance from the TechSec Observer newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What are the hardware requirements for Muse Spark Open?

It runs on two M5 Ultra Mac Studios, four Blackwell Ultra GPUs, eight H200s, or 300 GB GGUF setups.

How does it perform on coding benchmarks?

It scores 76.4 on Terminal-Bench 2.1, behind Kimi K3, GLM 5.3, and Qwen3.8.

What is the context window size?

The model supports a 1M-token context window.