Meta Releases Muse Spark Open for Local AI Work on Macs
The 816-billion-parameter open-weight model can run locally across two high-end Mac Studios with about 1 terabyte of combined unified memory.
Topics
News
- Meta Releases Muse Spark Open for Local AI Work on Macs
- Anthropic Gives METR Employee-Level Access to Review AI Risks
- Google Moves Links Higher in AI Search as Publishers Face Traffic Pressure
- CERT-In Says AI Agents Quietly Breached Indian Systems
- AI Dispatch | India Gives Diplomacy an AI Face
- Cisco Warns Critical Email Gateway Flaw Under Active Attack
Meta has released Muse Spark Open, an 816-billion-parameter open-weight AI model built for coding, multimodal work and agentic tasks, as the company expands its push to make more capable models available outside proprietary cloud services.
Meta called it the “most capable US open-weight model to date.” The model has 42 billion active parameters and supports a context window of up to 1 million tokens.
Muse Spark Open is designed for software engineering, multi-step reasoning and long-horizon agentic work. Meta has also optimized it to run locally on Apple hardware rather than relying solely on cloud infrastructure.
That still requires substantial computing power.
Meta said the full model can run across two Mac Studio systems equipped with Apple’s M5 Ultra chip using the MLX framework.
Each Mac Studio supports up to 512 GB of unified memory, giving the pair roughly 1 TB to hold the model.
The setup puts local deployment well beyond an ordinary Mac laptop or desktop and targets high-end workstation users.
Meta tested the model with Muse Code on several workflows that ran entirely on-device. In one test, the system made changes to Ice Cubes, an open-source Mastodon client, including fixing bugs, adding features and completing a software build.
The company also tested Muse Spark Open with DaVinci Resolve to review video footage frame by frame, identify useful sections and crop selected footage vertically.
Meta has positioned agentic performance as one of the model’s main strengths and compared it with other open-weight models across coding, computer-use and tool-use benchmarks.
Those results warrant some caution. Meta said some figures came from official leaderboards or rival model providers, while others came from its own testing. Its evaluation environment was also not specifically optimized for third-party models, meaning the comparisons may not reflect their best possible performance.
Muse Spark Open is available through several hosted AI platforms, while its model weights can be downloaded through services including Hugging Face, Google Cloud Model Garden, Amazon SageMaker AI and Microsoft Foundry Models.
