Technical Specifications and Capabilities
META PLATFORMS INC. has announced the release of Muse Spark Open, which the company describes as its most capable U.S. open-weight model to date. According to company statements, the model is specifically engineered for coding, agentic workflows, complex software engineering, and multi-step reasoning.
The model features a Mixture-of-Experts (MoE) architecture with 816 billion total parameters and 42 billion active parameters. It supports a context window of up to 1 million tokens and is natively multimodal, capable of processing text, image, video, and audio. To ensure safety, the company noted that the model underwent additional post-training evaluations conducted under Meta’s Superintelligence Scaling Framework.
Deployment and Optimization
The release includes several optimizations for hardware efficiency. The model ships with custom, auto-tunable Helion kernels in SGLang, which reportedly provide kernel speedups ranging from 1.03× to 3.95× across NVIDIA and AMD GPUs. On AMD MI350 hardware, the company claims a 3% end-to-end improvement.
Muse Spark Open is compatible with various inference engines, including vLLM, TensorRT-LLM, and llama.cpp, with optimizations available for NVIDIA, AMD, Intel, and Apple silicon. Notably, the model has been optimized for Apple M5 Ultra Mac Studios using the MLX framework.
The model weights are available for download through several major cloud providers, including AMAZON.COM INC (Amazon SageMaker AI), Google Cloud Model Garden, and Microsoft Foundry Models. It is also accessible via API platforms such as Groq, Together AI, and Fireworks AI. The weights are released under the Muse Model License, which includes limited commercial terms.
Community Reception and Performance Debates
Following the announcement, discussions on platforms such as Reddit have revealed a divide between the company’s claims and user-reported performance. While Meta positions the model as a leader in the open-weight space, some community members have questioned its competitive standing.
Unverified reports and user benchmarks shared on social media suggest that the model may underperform compared to recent releases from Chinese competitors, such as Qwen or GLM. Some users noted that the model’s performance appeared lower than Meta’s previous iterations, specifically the 1.2 and 1.3 versions.
Technical discussions also highlighted concerns regarding the model’s reasoning efficiency. Some users observed that the model appeared to require significant computational overhead for its output, comparing its performance unfavorably to smaller, more efficient “flash” models. Additionally, some developers expressed confusion regarding the strategic decision to release this specific version given the performance of existing open-source alternatives.