Muse Spark Open, the Most Capable US Open‑Weight Model to Date

Meta released Muse Spark Open, an 816B-total-parameter, 42B-active open-weight multimodal mixture-of-experts model with a 1M-token context window, tuned for agentic coding, long-horizon tasks and multi-step reasoning. The weights are available day one across Groq, Fireworks, Together, vLLM, SGLang, TensorRT-LLM and llama.cpp, and the model can run locally on two Apple M5 Ultra Mac Studios via MLX. Meta says the 42B active-parameter count and on-prem viability give teams a path to lower per-token agent costs and to keep sensitive codebases off third-party endpoints.

Hacker News https://research.meta.ai/blog/muse-spark-open Muse Spark Open, the Most Capable US Open‑Weight Model to Date Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. Meta released an 816B-parameter 42B active open-weight multimodal MoE with a 1M-token context, tuned for agentic coding and long-horizon tasks, and it can run locally on two Apple M5 Ultra Mac Studios via MLX. It's live day-one across the major inference providers Groq, Fireworks, Together, vLLM, SGLang, TensorRT-LLM, llama.cpp , so you can benchmark it against your current frontier API immediately—and the sparse active-param count plus on-prem viability means you now have a credible path to cut per-token agent costs and keep sensitive codebases off third-party endpoints. Meta released Muse Spark Open’s weights: an 816B-total-parameter, 42B-active natively multimodal model with a 1M-token context window aimed at coding, long-horizon agentic work, and multi-step reasoning. The practical shift is deployment flexibility: teams can run it via major API platforms, download weights from cloud/model hubs, or run it locally on two Apple M5 Ultra Mac Studios using MLX and other inference stacks, improving control over customization and sensitive workloads. AI vs. AI Debate “The summary incorrectly groups inference engines like vLLM, SGLang, TensorRT-LLM, and llama.cpp as “inference providers” and overstates cost savings as a consequence not established in the excerpt.” “The grouping reflects practical deployment reality—these engines are how you access the model—and the sparse 42B-active-parameter count is a well-established, direct driver of lower per-token compute cost, not an unsupported leap.”