Muse Spark Open

816B OPEN SOTA

Released Oct 15, 2026; Meta's most capable US open-weight model: 816B-total/42B-active MoE with native text, image, video, and audio input, 1M context, GPQA Diamond 89.0%, day-one inference on two M5 Ultra Mac Studios, and hosted APIs from $1.25/$4.25 per 1M tokens.

Context

1,000,000 tokens

Architecture

816B Total / 42B Active MoE (256 experts)

Official API (1M)

$1.25 in / $4.25 out (Open)+ Weights Free to Self-Host

License

Muse Model License

Published benchmarks · lab-reported

Vendor-published figures, not independently reproduced. Sources in Verification below.

TERMINAL-BENCH

76.4% (Terminal-Bench 2.1)

GPQA DIAMOND

89.0%

Quickstart

curl -s https://modelregistry.tirup.in/api/cli?model=muse-spark-open

OFFICIAL ANNOUNCEMENT WEIGHTS / REPOSITORY API DOCS

BACK TO LEDGER

Sources

Muse Spark Open announcement · Meta
Muse Spark Open model card · Meta
Muse Spark Open model page and benchmarks · Meta
Muse Spark Open listing and pricing · OpenRouter
Meta launches an open-weight AI model that runs on Macs · TechCrunch

Report an error in this record