Muse Spark 1.1
Released Jul 9, 2026 · knowledge cutoff unpublished
- Status
- Superseded
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 128K
- Speed (xhigh effort)
- 217 tok/s · 2.89s to first answer token
- Price ($/MTok in / out)
- $1.25 / $4.25
- Cost per task (xhigh effort)
- $0.26
- Open weights
- No
Benchmarks5 of 10 reported
GPQA Diamond
SWE-bench Pro
Terminal-Bench 2.1
Humanity's Last Exam
LMArena Elo
MMLU-Proretired
SWE-bench Verified
AIME (latest)retired
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Handles a multi-step browser task without hand-holding — tool and computer use are the point
- Takes text, images, video, PDFs and audio in a single prompt
- Good at creative and tone-shifting work: drafting, editing, adapting for an audience
Weaknesses
- Long-horizon agentic work still drifts where the leaders hold together
- Closed and metered, a departure from Meta's open-weights history
Second Meta Superintelligence Labs release, an upgrade to the original Muse Spark (April 2026) and Meta's first paid API model. HLE 62.1% (with tools) per Meta/BenchLM; Artificial Analysis separately reports a 40% → 45% no-tools HLE gain over the original Muse Spark. SWE-bench Pro 61.5% reported; no SWE-bench Verified figure published. GPQA Diamond 89.8% and LMArena Elo 1495 ('Muse Spark 1.1' text-arena listing, arena.ai) are third-party. ARC-AGI-2, MMLU-Pro, and AIME not published for this release. Knowledge cutoff not disclosed. Terminal-Bench 2.1 76.2% is Meta's own figure from the 2026-08-05 Muse Spark 1.2 launch chart, run in the generic mini-swe-agent harness (transcribed consistently by three outlets; the chart image itself was not readable this run), so it is not directly comparable with 1.2's 82.9% inside Muse Code.
For developers
API model strings
- meta
muse-spark-1.1
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor