Meta releases Muse Spark Open — a claimed "most capable US open-weight model"
From the AI digest for 2026-10-15
What happened: ~1:08 PM ET Oct 15, Meta AI (relayed by Yann LeCun): "Today we're releasing Muse Spark Open, the most capable US open-weight model to date. Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning." Key model specs on the release: 816B total / 42B active parameters, a 1M token context window, and natively multimodal across text, image, video, audio.
What's genuinely new: A development on the "Meta bets on Muse consumer agent" thread — a model-layer strand beside the Oct 12 Muse Wonder/Wonder Code launch (Wonder being the model intended to power the Muse agent): Spark carries the open-weight flag, the biggest-parameter disclosure Meta has had on this record, and a "most capable US open-weight model" superlative asserted on the company's own channel, with no benchmark, license terms, or weights-availability detail in the window. The 42B-active/816B-total sparse shape and 1M context put it in the frontier-adjacent open-weight lane where the record's comparisons have sat (GPT-6-Astra, open-weight releases from other labs) — but everything here is the announcement's say-so.
Watch: License and where the weights actually land; independent benchmarks against the "most capable US open-weight" claim; the role Spark Open plays in the Muse stack (training substrate or inference upgrade for Muse-on-Wonder); download/usage traction.