← Back to directory
Meta AI
Muse Spark 1.2
Released Aug 5, 2026 · knowledge cutoff unpublished
- Status
- Frontier
- Location
- United States
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 131K
- Speed (xhigh effort)
- 165 tok/s · 26s to first answer token
- Price ($/MTok in / out)
- $1.25 / $4.25
- Cost per task (xhigh effort)
- $0.40
- Open weights
- No
Benchmarks1 of 8 reported
Terminal-Bench 2.1
82.9%
MMLU-Pro
—
GPQA Diamond
—
SWE-bench Verified
—
AIME (latest)
—
Humanity's Last Exam
—
LMArena Elo
—
ARC-AGI-2
—
Strengths & weaknesses
Strengths
- Built for multi-agent workflows — plans and delegates as a main agent, or runs in parallel as a subagent, with structured output and parallel function calling across coding harnesses
- Coding-focused retrain shows in whole-repository work — holds dependency graphs and legacy code together across the long context window better than 1.1 did
- Ships alongside Muse Code, a purpose-built terminal coding agent co-trained with the model rather than bolted on after release
Weaknesses
- Reasoning can't be switched off, and the coding-focused retrain cost latency — time to first token runs roughly nine times longer than 1.1, trading speed for the index gain
- Agentic knowledge work — the gap Meta itself flagged in 1.1 — narrows but still trails the frontier
Terminal-Bench 2.1 score of 82.9% was measured running inside Muse Code, the paired coding agent Meta shipped alongside it, rather than the generic harness used for 1.1's SWE-bench Pro figure, so the two releases aren't directly comparable; GPQA Diamond, MMLU-Pro, AIME, HLE, LMArena Elo, and ARC-AGI-2 were not published for this release, and knowledge cutoff remains undisclosed. Meta's own announcement calls Muse Spark 1.1 'its predecessor,' the basis for predecessorId.
For developers
API model strings
- meta
muse-spark-1.2
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
Replaces Muse Spark 1.1