# Benchmark methodology: https://research.meta.ai/static/muse-spark-1-3-multimodal-evaluation-methodology
# Scores: https://research.meta.ai/articles/introducing-muse-1-3/benchmarks/benchmark-scorecard-v6.webp
# Sources:
# https://research.meta.ai/blog/introducing-muse-spark-1-3
# https://dev.meta.ai/docs/models
# https://openrouter.ai/meta/muse-spark-1.3 (OpenRouter Meta-hosted catalog snapshot, 2026-09-02)

name = "Muse Spark 1.3"
description = "Muse Spark 1.3 is a multimodal reasoning model from Meta for long-running agentic, multi-agent, and coding workflows. It improves long-horizon agent collaboration, instruction following, and coding efficiency relative to Muse Spark 1.2."
family = "muse"
release_date = "2026-09-02"
last_updated = "2026-09-02"
attachment = true
reasoning = true
temperature = true
tool_call = true
structured_output = true
open_weights = false

[limit]
context = 1_048_576
output = 131_072

[modalities]
input = ["text", "image", "video", "pdf", "audio"]
output = ["text"]

[[benchmarks]]
name = "GDPval-AA"
score = 1754
metric = "Elo"
variant = "max effort"
version = "2"
harness = "Stirrup"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "JobBench"
score = 64.9
metric = "mean rubric score"
variant = "max effort"
harness = "OpenCode"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "OSWorld"
score = 66.9
metric = "mean partial score"
variant = "max effort"
version = "2.0"
dataset = "08.08"
harness = "Meta internal evaluation framework"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "OSWorld"
score = 32.0
metric = "binary completion rate"
variant = "max effort"
version = "2.0"
dataset = "08.08"
harness = "Meta internal evaluation framework"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "DeepSearchQA"
score = 90.3
metric = "F1"
variant = "max effort"
harness = "Meta browser harness"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "Agentic IF Index"
score = 57.8
metric = "score"
variant = "max effort"
dataset = "Meta internal"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "AutomationBench"
score = 49.6
metric = "pass@1"
variant = "max effort"
dataset = "public v3 task set"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "MRCR"
score = 98.5
metric = "mean sequence-match ratio"
variant = "max effort"
version = "2"
dataset = "8-needle, 256K-512K"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "MRCR"
score = 98.1
metric = "mean sequence-match ratio"
variant = "max effort"
version = "2"
dataset = "8-needle, 512K-1M"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "DeepSWE"
score = 75.4
metric = "pass@1"
variant = "max effort"
version = "1.1"
harness = "mini-swe-agent"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "SWE-Atlas Codebase QnA"
score = 59.4
metric = "pass@1"
variant = "max effort"
harness = "mini-swe-agent"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"

[[benchmarks]]
name = "Terminal-Bench"
score = 88.8
metric = "pass@1"
variant = "max effort"
version = "2.1"
harness = "Muse Code"
source = "https://research.meta.ai/blog/introducing-muse-spark-1-3"
date = "2026-09-02"