Muse Spark
Released Apr 8, 2026 · knowledge cutoff unpublished
- Status
- Superseded
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 262K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- No
Benchmarks6 of 10 reported
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Humanity's Last Exam
LMArena Elo
ARC-AGI-2
MMLU-Proretired
Terminal-Bench 2.1
AIME (latest)retired
GDPval-AA v2
Strengths & weaknesses
Strengths
- Health and medical reasoning is its clear specialism, not a side capability
- Free inside Meta's consumer apps, so its real user base dwarfs its developer footprint
- Deliberate 'contemplating' mode for questions worth extra time
Weaknesses
- Weak on abstract novel reasoning relative to how strong it is on domain knowledge
- No public API at launch — usable only through Meta's own products
Replaced Llama as Meta's flagship line. Closed hosted model — no public API pricing or max-output figure. SWE-bench 77.4 lab-reported; HLE 58.4 is third-party (LLM-Stats), with tools in Contemplating mode; ARC-AGI-2 42.5 in Contemplating mode. SWE-Bench Pro (public) 52.4 is Meta's own figure for Muse Spark (Thinking) in its Muse Spark Eval Methodology table (ai.meta.com/static-resource/muse-spark-eval-methodology; Meta's scaffold, averaged over 4 attempts); Scale's public leaderboard lists "Muse Spark" higher at 55.00% — lab figure used, per house rules.
For developers
API model strings
Not researched
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor