Meta’s Muse Spark Open Enters Arena.ai for Public Testing
Key Takeaways
- •Meta announced Muse Spark Open on October 15, 2026, marking the first open-weight release in its previously proprietary Muse Spark series.
- •The multimodal model has 816 billion total parameters, 42 billion active parameters, a 1M-token context window, and can process images, video, and audio.
- •Developers can evaluate the model with live performance scores in Arena.ai's Agent Arena and Code Arena, providing a third-party check on Meta's benchmark claims.
- •Access options include API providers such Baseten and Fireworks AI, while model weights can be downloaded through Amazon SageMaker and Google Cloud Model Garden.
- •Meta optimized the model for Apple silicon via the MLX framework, citing a reference setup of two M5 Ultra Mac Studios with 1TB of unified memory and safety post-training that reportedly allows operation without external guardrails.
Meta’s latest artificial intelligence model is now being tested by other developers through Arena.ai.
Muse Spark Open, the first open-weight release in Meta’s Muse Spark series, is available in Arena.ai’s Agent Arena and Code Arena. Users can evaluate the model on complex tasks and see live performance scores. Meta announced the model on October 15, 2026, and Arena.ai confirmed its integration and live testing that day. The announcement was reported by CryptoBriefing.
Model capabilities
Muse Spark Open is a multimodal model with 816 billion total parameters, including 42 billion active parameters. It has a 1M-token context window and can process images, video, and audio.
Meta is positioning the model for long-horizon agentic workflows, complex coding, and multi-step reasoning. For coding tasks, Meta says Muse Spark Open can analyze entire software repositories and handle complicated refactors spanning multiple files — use cases that draw on the model’s technical profile, pairing the 1M-token context window with native image, video, and audio processing.
Access and infrastructure
Muse Spark Open became available on October 15, 2026, through various API platforms and cloud marketplaces. API access includes providers such as Baseten and Fireworks AI, among others. Organizations seeking the model weights can download them through Amazon SageMaker and Google Cloud Model Garden.
Meta says Muse Spark Open can run on users’ own infrastructure and has been specifically optimized for Apple silicon through the MLX framework. Its reference configuration consists of two M5 Ultra Mac Studios connected as a supercomputer with 1TB of unified memory — a reference setup that puts self-hosting on workstation-class hardware rather than data-center clusters.
Meta also says the model underwent safety post-training. According to the company, this allows Muse Spark Open to operate independently without additional external guardrails — a detail that bears directly on organizations weighing a self-hosted deployment.
Testing through Arena.ai
Muse Spark Open builds on Meta’s earlier Muse Spark models, which were proprietary. The release is the series’ first move to open weights.
Meta Superintelligence Labs claims that Muse Spark Open matches frontier closed models on agentic coding benchmarks. Agent Arena and Code Arena evaluate models on practical tasks involving coding, web search, and multimedia processing, while live scoring shows how they perform — giving outside developers a running, third-party check on those claims as testing continues.
