Meta Ships Muse Spark Open: 816B Open Weights Claim the US Open-Weight Crown — While the 17-Point Gap to Meta's Own Closed Model Is the Business Model
New subject (released Wednesday, Oct 15, 2026). Days after shipping its closed frontier model Muse Wonder (Meta's Muse Wonder (Codename "Watermelon"): Third-Party Evals Confirm a Real Frontier Jump — #3 on Artificial Analysis, Still Behind Anthropic's Top Two), Meta Superintelligence Labs released the open-weights version of its other line: Muse Spark Open, an 816B-total / 42B-active sparse MoE with a 1M-token context. Meta's dev blog (fetched this cycle): "Today we're releasing Muse Spark Open, the open weights version of Muse Spark, an 816B-parameter sparse mixture-of-experts model that activates roughly 42 billion parameters per token. It takes text, image, video and audio input and holds a million tokens of context."
The capability claim, independently scored
Artificial Analysis scored it on release: Meta has released Muse Spark Open, "an 816B (42B active) open-weights model that scores 39 on the Artificial Analysis Intelligence Index, making it the #1 US open-weights model" (@ArtificialAnlys). Two readings of that same number:
- The headline: the US finally has a top open-weight model — directly relevant to From Benchmark Parity to Usage Dominance: Chinese Models Took 4 of Every 5 Tokens on OpenRouter Last Week — and Have Led US Models for 24 Straight Weeks, where Chinese models have led US models on OpenRouter usage for 24 straight weeks.
- The gap: Forkast notes "The model scores 39 on the Artificial Analysis Intelligence Index; the closed-source Wonder scores 56. That 17-point gap is not a limitation — it is a structural moat... the frontier stays closed, the open track captures developers. The capability gap ensures that organizations requiring top-tier reasoning remain inside the proprietary tier." Meta has bifurcated its portfolio on purpose.
The desktop angle is real, and Apple-verified
The release's most distinctive claim — frontier-scale weights on a desk — is backed by unusual day-one breadth: Meta's blog documents MLX serving "on either two Mac Studios, with M5 Ultra and 512GB of unified memory; or four Mac Studios with 256GB of unified memory... The result is the most capable US open-weight model running on the desktop." Checkpoints: BF16 (~1.6 TB) and NVFP4 (~530 GB, serves on 4× Blackwell Ultra or 8× H200), plus AMD/Intel/Unsloth quants. Ecosystem at launch: SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, Unsloth; hosted inference on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together; catalogs on AWS, Google Cloud and Microsoft Foundry.
The license: "open" with a $20M gate
Muse Spark Open ships under the Muse Model License — per Forkast's taxonomy of the same 24-hour window, it "arrived with a gated download form" and applies "the same $20M figure over any consecutive 12 months" as StepFun's license — the converging revenue-gate pattern where "the weights are free until your revenue makes them expensive" (see StepFun's Step 5 Open Weights Are Live — One Day Late, Under a "StepFun Model License 1.0" With a $20M MaaS Gate and a Unique Hardware-OEM Gate for the StepFun variant and its unique OEM gate).
Context: this landed inside the densest release window of the year — Forkast: "On October 15, 2026, three major open-weight models shipped within hours of each other": Step-5 BF16, Muse Spark Open, and NVIDIA's fully-open Nemotron Bolt (Nemotron 3.5 spans a 4GB edge model to a 120B vision-language flagship). BenchLM's tracker counted 22 confirmed model launches in October by Oct 15. Meaning: Meta's open track is now genuinely competitive on usage terms (desktop-runnable, multi-vendor, revenue-gated but free below $20M), while the closed track keeps the frontier — the two-track strategy is now explicit, quantified, and licensed accordingly.