Meta dropped a bombshell in the open-weight AI race on October 15th: Muse Spark Open, a massive 816-billion-parameter mixture-of-experts model released freely under the Muse Model License. It's now the most capable US open-weight model available — and the hardware requirements are genuinely staggering.
What It Is
Muse Spark Open is the open-weight counterpart to Meta's frontier Muse Spark model, released to Hugging Face and major cloud marketplaces including Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry. Despite its size, the MoE architecture activates only about 42 billion parameters per token — making inference significantly cheaper than its raw parameter count implies.
The model takes text, images, video, and audio as input, and comes with a 1-million-token context window. Meta has specifically optimized it for long-horizon coding tasks, complex software engineering challenges, and multi-step agentic reasoning — positioning it directly against Chinese open-weight heavyweights like GLM 5.3 and Kimi K3.
The Mac Angle — and the Catch
Meta has made a deliberate push to make Muse Spark Open run locally on Apple Silicon. TechCrunch reports the model is optimized for two M5 Ultra Mac Studios running in tandem. The catch? You'll need at least 1 terabyte of RAM. Engadget estimates that a four-machine setup using 256GB M5 Ultra Mac Studios would run roughly $38,000. This isn't a model you'll be running on your laptop.
For those without a six-figure hardware budget, API access is available through Meta's developer platform, and the model is compatible with popular inference frameworks including SGLang, vLLM, TensorRT-LLM, and llama.cpp.
Benchmarks: Impressive, with Asterisks
Meta's own benchmarks show Muse Spark Open outperforming GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. MarkTechPost's independent analysis paints a more nuanced picture: the model lands mid-pack among large open models on pure coding benchmarks and trails Kimi K3 on Terminal-Bench 2.1. As always, treat vendor-provided benchmarks with appropriate skepticism.
Why It Matters
The open-weight frontier has been largely dominated by Chinese labs for months. Muse Spark Open shifts that balance, giving researchers, enterprises, and hobbyists (with deep pockets) access to a genuinely frontier-class model they can fine-tune, audit, and self-host. For the agentic AI use cases that are increasingly driving commercial value, having a powerful open alternative to closed APIs is significant.
The release also continues Meta's strategy of using open weights as a competitive lever — building ecosystem goodwill and developer lock-in while pressuring closed competitors on price and accessibility.