Meta open-sourced 816 billion parameters. The model is real. The license has a catch.

Meta released the weights for Muse Spark Open on October 15. The headline specs: 816 billion total parameters, 42 billion active at inference, a 1M-token context window, multimodal inputs (text, image, video, audio), and text output. Weights are on Hugging Face at meta-models/Muse-Spark-Open-816B-A42B. The hosted API is live on OpenRouter and Together AI.

Meta's own benchmark table shows MCP-Atlas at 89.2% — a tool-use benchmark — and Terminal-Bench 2.1 at 76.4%. Those numbers are Meta's. No independent verification at launch. The license is called the Muse Model License, and it is the thing worth understanding before you build on this.

Source spread

• Meta AI Developer Blog — Build with Muse Spark Open — hype. Official launch post. Source for specs, benchmarks, and license summary.
• PyTorch Blog — Muse Spark is now an open-weights model — builder. Technical integration details and architecture overview.
• The New Stack — Meta ships Muse Spark Open — skeptic. Flags the Muse Model License as "semi-open" and compares it to Apache 2.0. Best critical take in the coverage.
• technology.org — Meta Muse Spark Open with Mac Studio — builder. Details the 2×M5 Ultra deployment path and MLX quantization options.

Pros & cons

What's real:

• The scale is competitive. 816B total parameters with 42B active is a serious MoE architecture. For comparison: Kimi K3 is 2.8T total (larger), but active-parameter efficiency matters more in practice. At 42B active, this is fast where it counts.
• The multimodal input story is good. Text, image, video, and audio in a single model, under a single call. Most competitors are still stitching this together from multiple specialized models. Builders who need unified multimodal pipelines don't have many open-weight alternatives at this scale.
• Hardware is coming down. 700GB of GPU memory in NVFP4 quantization is expensive but reachable for well-funded orgs. Community 2-bit quantization gets it to ~300GB. And two M5 Ultra Mac Studios — each with 512GB unified memory — can run it via MLX. That's a weird but real on-premise path for teams that need it.
• Additional safety post-training. Meta included extra safety post-training in the open-weights release to make the model more robust without system guardrails. That matters for builders deploying without a full safety stack on top.

What deserves scrutiny:

• The benchmarks are first-party. MCP-Atlas 89.2% and Terminal-Bench 2.1 76.4% are Meta's own reported numbers. First-party benchmark claims have a consistent track record of being optimistic. Trust them provisionally; wait for independent evaluation.
• The Muse Model License is not Apache 2.0. The New Stack correctly calls this semi-open. You can use it, fine-tune it, distill it, redistribute it. What you cannot do is ship a product built on it without displaying a "Built with Muse Spark Open" label. That's not a minor footnote. It's a licensing constraint that makes this model a distribution channel for Meta's brand in everything built on top of it.
• BF16 requires 2TB. If you want the full-precision model without quantization, you need 2 terabytes of GPU memory. That's not a commodity setup. Most teams will be running quantized versions, with the attendant quality tradeoffs.

What builders need to know

• The license requires attribution in shipping products. Read the Muse Model License before building anything customer-facing. Any shipped product needs a visible "Built with Muse Spark Open" label. Internal tools are fine; consumer products need the label.
• Distillation is permitted. You can distill this model into a smaller one and ship the distilled version. Check whether the attribution requirement follows through to the distilled model — the license language matters here.
• Start with hosted API, not local weights. The model is on OpenRouter and Together AI with no infrastructure setup required. Evaluate capability there before committing to GPU provisioning.
• MCP-Atlas 89.2% matters more than the headline number. For agentic work — tool use, multi-step task completion — MCP-Atlas is the benchmark to watch. 89.2% is Meta's claim; independent scores will follow. That's the number to track.
• Quantization community will be active. The open-weights release means community quantizations at 4-bit and 2-bit are already in progress. GGUF and MLX builds will be faster and smaller than the official NVFP4 release.

Further reading

• Meta AI Developer Blog — Build with Muse Spark Open — official launch post with specs and benchmarks
• PyTorch Blog — Muse Spark is now an open-weights model — technical integration
• The New Stack — Meta ships Muse Spark Open, its answer to China's best open-weight models — best critical take on the license
• technology.org — Meta Muse Spark Open: 816B open-weight with Mac Studio path — hardware deployment details
• HuggingFace — meta-models/Muse-Spark-Open-816B-A42B — weights and model card