AI model reference · 2026

Muse Spark 1.1

MetaModel details, evidence, and API facts

Muse Spark 1.1 is a multimodal reasoning model from Meta, built for agentic tasks.

Text generation

Image input

Reasoning

Provider endpoint accepts tool parameters

Try Muse Spark 1.1 Compare with another model

Context window: 1,048,576 tokens
Maximum output: 131,072 tokens
Base provider API input: $1.25 / 1M tokens; Base provider API rate
Published evidence: 12 benchmarks, 2 Arena results

Overview

About Muse Spark 1.1

A concise catalog overview. Technical limits and published evaluation evidence are listed separately below.

Muse Spark 1.1 is a multimodal reasoning model from Meta, built for agentic tasks.

text

multimodal

reasoning

Published evaluations

Benchmark results

Only results mapped to this exact model and backed by a named source appear here. Different evaluation protocols are not treated as interchangeable.

Source-backed benchmark results for Muse Spark 1.1
Benchmark	Score	Evaluation details	Source
AutomationBench-AAA 657-task benchmark of multi-step work across simulated SaaS applications in six business domains. The Artificial Analysis protocol reports a guardrail-aware score and is distinct from both the public Zapier split and the unrelated dynamic AutoBench framework.	42.8%	Guardrail-aware scoreVersion: AutomationBench-AA 2026 · Split: 657 tasks across Finance, HR, Marketing, Operations, Sales, and Support · Harness: Artificial Analysis independent implementation · Attempts: One evaluated run per task · Tools: REST APIs for simulated SaaS applications · Evaluator: Programmatic end-state grader with guardrail checks · Reasoning effort: xhighMethodology	Artificial AnalysisIndependently reproduced
Finance Agent v2A Vals AI benchmark with 927 expert-reviewed questions modeling the work of entry-level financial analysts.	57.2%	Task scoreVersion: Finance Agent v2 · Split: 927 expert-reviewed questions · Harness: Shared six-tool harness · Evaluator: Partial Credit metric · Reasoning effort: xhigh	Vals AIBenchmark owner
JobBenchA professional tool-use benchmark with 65 tasks spanning 35 white-collar occupations.	54.7%	Task scoreVersion: 65 tasks across 35 occupations · Harness: OpenCode v1.14.18 · Evaluator: Grok 4.3 judge · Reasoning effort: xhigh	JobBenchBenchmark owner
MCP AtlasA scaled tool-use benchmark with 1,000 multi-step tasks across 36 real MCP servers and 220 tools.	88.1%±1.95	Tasks completedVersion: 1,000 tasks across 36 MCP servers and 220 tools · Reasoning effort: xhigh	Scale AIBenchmark owner
OSWorld-VerifiedA verified computer-use benchmark in which multimodal agents operate desktop applications and are graded from the resulting environment state.	80.8%	Mean task rewardVersion: OSWorld-Verified · Split: 361 tasks excluding Google Drive-dependent tasks · Attempts: 1 per task · Tools: GUI computer-control only; no shell · Reasoning effort: xhighProtocol note: 1920x1080 screenshots, 200-step cap, screenshot truncation enabledMethodology	MetaProvider reported
DeepSWE 1.1A long-horizon software-engineering benchmark with 113 original tasks graded by hand-written tests.	53.0%±3	Pass@1Version: 1.1 · Split: 113 tasks · Harness: bash-only mini-swe-agent fork · Attempts: 5 per task · Tools: No Internet · Reasoning effort: xhighMethodology	DataCurveBenchmark owner
SWE-Bench ProA contamination-resistant software-engineering benchmark with long-horizon tasks across multiple programming languages. Public and private splits are distinct protocols.	61.5%±3.1	ResolvedVersion: SWE-Bench Pro · Split: 731 public tasks · Harness: mini-swe-agent · Reasoning effort: xhighMethodology	Scale AIBenchmark owner
Terminal-Bench 2.1Version 2.1 of the benchmark for completing realistic tasks in terminal environments. Harness, resource limits, and attempt count are part of the protocol.	80.0%	Mean task successVersion: 2.1 · Split: 89 tasks · Harness: bash-tool-only agent · Attempts: 5 per task · Reasoning effort: xhighProtocol note: 6 CPU cores, 8GB RAMMethodology	MetaProvider reported
Humanity's Last ExamA 2,500-question expert-level benchmark spanning dozens of academic fields. Tool-assisted and no-tools results are separate protocols and must not be merged.	62.1%	AccuracyVersion: Full 2,500-question dataset · Tools: Bash and browser tools with source blocklist · Evaluator: GPT o3-mini judge · Reasoning effort: xhighMethodology	MetaProvider reported
LiveBenchA contamination-resistant benchmark refreshed on a fixed release cadence. Scores from different LiveBench releases must never be compared as the same protocol.	75.3%	Mean of category averagesVersion: LiveBench 2026-06-25 · Split: 2026-06-25 release, 23 tasks across seven categories · Reasoning effort: xhighProtocol note: Overall is the mean of category averages. This protocol is not comparable with the 2026-01-08 release.Methodology	LiveBenchBenchmark owner
BabyVisionA 388-question visual-understanding benchmark covering fine-grained discrimination, spatial perception, tracking, and pattern recognition.	76.3%	AccuracyVersion: 388 questions · Tools: Code execution · Evaluator: gpt-oss-120b at high reasoning effort · Reasoning effort: xhigh	MetaProvider reported
CharXiv-RThe chart-reasoning portion of CharXiv, evaluating visual and mathematical reasoning over scientific figures.	88.4%	AccuracyVersion: CharXiv validation set · Split: 1,000 chart-reasoning questions · Tools: Code execution · Evaluator: gpt-oss-120b at high reasoning effort · Reasoning effort: xhighMethodology	MetaProvider reported

Human preference

Arena results

Arena scores come from blind human preference votes. They are reported separately from task benchmarks and are not used as substitutes for missing benchmark results.

Arena results for Muse Spark 1.1
Category	Score	Rank	Votes	Status
Arena TextBlind human preference score for text responses.	1,491±4	#12	37,238	Published
Arena WebDevBlind human preference score for code and web development.	1,542±8	#39	8,070	Published

Technical reference

Specifications and API facts

Technical limits and reference provider prices come from the current catalog unless a reviewed external source is listed. Missing values stay marked as unavailable.

Model