Must read: Meta takes on OpenAI and Anthropic with Muse Wonder; Aims for personal superintelligence
Muse Spark Open: Performance and benchmark
Meta says the Muse Spark Open leverages Apple’s latest version of the MLX framework, which is optimised for the unified memory architecture of Apple silicon. Hence, the model is capable of running on a Mac Studio with M5 Ultra, allowing users to run powerful models on a desktop.
In terms of benchmarks, the Muse Spark Open brings competitive performance across developer-focused and multimodal evaluations, highlighting its strengths in agentic coding and terminal-based problem solving.
The model scored 76.4 on Terminal-bench 2.1, highlighting its software engineering benchmarks. Across broad knowledge and academic reasoning evaluations such as SciCode and AA-Omniscience, the model is positioning itself near top-tier open-weight counterparts.
Must read: Meta Muse AI is learning more about you than you may realise: Here’s what it tracks
How can Mac Studio users benefit from Muse Spark One?
Meta says Muse Spark One benefits users and developers in several key ways, as it is the first open-weight model in its class capable of running directly on desktop. With the MLX framework, users can seamlessly run a powerful superintelligence-class model locally on their own devices.
It is built specifically for coding and agentic work, and Meta claims that it can seamlessly manage multi-step reasoning, complex software engineering, and extensive projects end-to-end. With its 1-million-token context window, users can process vast amounts of data, large codebases, or documentation simultaneously without losing track of details.
Muse Spark Open availability
Muse Spark Open is available via Hugging Face and enterprise cloud platforms, such as Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. It is also available via API across Together AI, Fireworks AI, Groq, OpenRouter, Baseten, Modal, and Nebius.