# Meta Muse Spark Open: 816B Agent Model, One License Gotcha > Source: > Published: 2026-10-16 10:10:30+00:00 Meta opened the weights on its Muse Spark frontier model on October 15, 2026. The download is live on Hugging Face. 816 billion total parameters, 42 billion active via mixture-of-experts routing, 1 million token context window, native multimodal. On agent and tool-calling benchmarks, Muse Spark Open leads every open-weight model tested. On coding tasks it trails China’s best by 8–10 points. And if you plan to resell API access at scale, there’s a $20 million revenue clause buried in the license you need to read before you ship. ## What Shipped Muse Spark Open runs a 256-expert MoE transformer with only 8 experts active per token — meaning the compute profile at inference is closer to a 42B dense model than an 816B one. The 1M token context window is included by default, not a premium tier. Multimodal inputs (text, image, video, audio) are supported through a 1.8B perception encoder baked into the architecture. The weights are available in BF16 and NVFP4 quantized format on [Hugging Face](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B). Seven inference providers offer hosted access at launch: Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, OpenRouter, and Together AI. If you want to try it today without touching a GPU, OpenRouter and Together AI are the fastest path. ## Where It Leads, Where It Doesn’t Vendor-reported benchmarks at xhigh reasoning effort tell two different stories: | Benchmark | Muse Spark Open | GLM 5.3 | Kimi K3 | |---|---|---|---| | MCP-Atlas (tool-calling) | **89.2** | 84.2 | 82.3 | | Terminal-Bench 2.1 (coding) | 76.4 | 83.9 | **85.0** | | DeepSWE 1.1 (coding agents) | 60 | 67 | **68** | For multi-MCP-server deployments and long-horizon agent workflows, Muse Spark Open leads by a clear margin. For pure coding — the dominant local inference use case — it trails GLM 5.3 and Kimi K3 by 8–10 points on Terminal-Bench and about 8 points on DeepSWE. That gap matters if you’re evaluating this as a backend for a coding agent. If your workload is tool routing, agentic planning, or enterprise automation, the math looks different. Community consensus on [Hacker News](https://news.ycombinator.com/item?id=50107379) aligned quickly: Muse Spark Open fills the niche of companies that can’t or won’t run Chinese models but need open-weight inference. That’s a real and growing category given US enterprise compliance requirements around model provenance. ## The License Meta’s Muse Model License follows the same pattern as recent Llama releases — open weights, proprietary license. The practical breakdown: - **Free with no agreement required:** individual developers, researchers, companies building products that use the model internally or as a feature of their own service - **Requires a separate Meta agreement:** any company hosting the model as an API service for third parties if its aggregate gross revenue exceeds $20 million in any 12-month period - **Prohibited entirely:** military applications, critical infrastructure, malware generation, unauthorized professional services - **Required everywhere:** “Built with Muse Spark Open” attribution in UI or documentation Most developers building applications are well inside the free tier. If you’re an AI infrastructure company that sells model access and already clears $20M annually, the license is relevant now — not after you’ve shipped. The full terms are on the [Hugging Face model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B). ## How to Run It The hardware requirements are substantial. The NVFP4 quantized version needs 700GB of VRAM — four Blackwell Ultra GPUs or eight H200s. Full BF16 requires 2TB. For most teams, inference APIs are the right answer. If you have the GPU budget or want to self-host, vLLM works today: ``` vllm serve meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --tensor-parallel-size 4 --enable-expert-parallel --max-model-len 1048576 ``` The Mac path is genuinely interesting: two M5 Ultra Mac Studios running MLX carry the model. At roughly $10,000 in hardware, it’s not a personal laptop setup, but it’s the most accessible frontier-class local inference option that’s ever existed. Meta timed the release with [Apple’s M5 Ultra Mac Studio availability](https://www.macrumors.com/2026/10/15/meta-muse-spark-open-mac-studio/) deliberately. Unsloth’s 2-bit GGUF quantization brings the combined RAM and VRAM requirement down to around 300GB — workable for high-memory workstations for testing. ## Bottom Line Muse Spark Open is the best US-origin open-weight model for agentic and tool-calling workloads. Read the license before you build a hosted API service on it. If you’re evaluating it as a coding model, run the benchmarks against your own workload — the 8-point gap on Terminal-Bench compared to Kimi K3 is real and likely shows up in practice. For everyone else, Groq and Fireworks AI have it hosted today. Start there.