Wait Which Model?
← Back to directory
Meta AI

Muse Spark Open

Released Oct 15, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
United States
Type
LLM
Modality
Multimodal
Context window
1M
Max output
—
Speed (xhigh effort)
141 tok/s · 0.80s to first answer token
Price ($/MTok in / out)
$1.25 / $4.25
Cost per task (xhigh effort)
$0.84
Open weights
Yes
News1 story
Strengths & weaknesses

Strengths

  • Meta's open weights at a size that runs on two linked M5 Ultra Mac Studios, not only on a datacentre rack
  • Takes audio and video natively as well as images — still rare among open-weight models
  • Gets to an answer in fewer output tokens than the median model on Artificial Analysis' index despite defaulting to its heaviest reasoning setting

Weaknesses

  • The Muse Model License requires a separate agreement with Meta once a hosted service passes $20M in annual revenue, plus mandatory 'Built with Muse Spark Open' attribution — not a drop-in for a permissive licence
  • Defaults to xhigh reasoning and rejects effort values outside low–xhigh, so quick tasks need effort set explicitly
  • Every launch figure was run by Meta in its own harness and the detailed evaluation report is still to come — independent reception is thin

816B total / 42B active MoE with Meta's Perception Encoder (text, image, video and audio in; text out), released as open weights on 2026-10-15 with an extra round of safety post-training over the closed Muse Spark; NVFP4 needs about 700 GB of VRAM and BF16 about 2 TB. Benchmarks are Meta's model-card figures at xhigh: GPQA Diamond 89.0, Terminal-Bench 2.1 (Terminus 2) 76.4 and GDPval-AA v2 1401 (lab-reported); hle is null because the card reports 'HLE-Diamond' (no tools, 26.1), a subset, alongside HLE with tools 61.0. Also reported: MCP-Atlas 89.2, OSWorld Verified 80.4, Toolathlon Verified 72.8, SciCode 58.5, DeepSWE 1.1 60. Meta does not serve it on its own API; $1.25/$4.25 is OpenRouter's listing and Artificial Analysis' provider median, with hosts including Together, Fireworks and Groq. costPerTask and speed are AA's xhigh figures (Intelligence Index 39, $0.84/task, 141.0 tok/s, 0.80s time to first token as AA reports it); AA lists no other effort variant. Licence classified restricted for the $20M Model-as-a-Service revenue trigger and attribution clause; commercial use is otherwise free. maxOutput and knowledge cutoff are unpublished, and the card names no model it replaces, so predecessorId is null.

Benchmarks3 of 10 reported

GPQA Diamond

89%

Terminal-Bench 2.1

76.4%

GDPval-AA v2

1401

MMLU-Proretired

—

SWE-bench Verified

—

SWE-bench Pro

—

AIME (latest)retired

—

Humanity's Last Exam

—

LMArena Elo

—

ARC-AGI-2

—
For developers

API model strings

  • openroutermeta/muse-spark-open

Licence

Muse Model License · Restricted · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor