- CharXiv Reasoning (Figure Understanding) 86.4
- MMMU Pro (Multimodal Understanding) 80.4
- ERQA (Embodied Reasoning) 64.7
- SimpleVQA (Visual Factuality) 71.3
- ScreenSpot Pro (Screenshot Localization - With Python) 84.1
- ZeroBench (Multi-Step Visual Reasoning (pass@5) - With Python) 33
- Best of Humanity's Last Exam (No Tools) / Humanity's Last Exam (With Tools) 50.4
- ARC AGI 2 (Abstract Reasoning Puzzles (Public)) 42.5
- GPQA Diamond (PhD Level Reasoning) 89.5
- LiveCodeBench Pro (Competitive Coding) 80
- HealthBench Hard (Open-Ended Health Queries) 42.8
- MedXpertQA (Text) 52.6
- MedXpertQA (MM) 78.4
- DeepSearchQA (Agentic Search) 74.8
- SWE-Bench Verified (Agentic Coding) 77.4
- SWE-Bench Pro (Diverse Agentic Coding) 52.4
- Terminal-Bench 2.0 (Agentic Terminal Coding) 59
- τ²-Bench Telecom (Agentic Tool Use (Artificial Analysis)) 91.5
- GDPval-AA Elo (Office Tasks (Artificial Analysis)) 1444
- FrontierScience Research, Contemplating mode, 38% 38
Muse Spark
Muse Spark ranks first on 3 of the 73 benchmarks with five or more models scored.
101
Benchmarks with a score
3
Ranked first, out of five or more models
13
In the top three
Meta Apr 2026 Muse Spark
Closed weights Flagship
153.1
#53 of 256
Epoch Capabilities Index
Self-Reported Index 153.5 #71 of 404
- Humanity's Last Exam 40.6% #7 /47
- SWE-Bench Pro 55% #3 /24
- MMMU-Pro 80.4% #17 /90
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LMArena (Elo) | 1,489 | #15 of 335 | Official leaderboard | |
| MultiChallenge | 75.5% | #3 of 42 | Official leaderboard | |
| MultiNRC | 59% | #5 of 38 | Official leaderboard | |
| SimpleQA Verified | 70.8% | #3 of 22 | Lab model card | |
| Humanity's Last Exam | 40.6% | #7 of 47 | Official leaderboard | |
| SimpleVQA | 71.3% | #9 of 30 | Lab model card | |
| Document Arena | 1,467 | #19 of 39 | Official leaderboard | |
| AbstentionBench | 59.1% | #1 of 2 | Lab model card | |
| HLE Calibration | 50.3 | #3 of 3 | Lab model card | |
| CharXiv Missing Images | 35% | #2 of 2 | Lab model card | |
| Meta Underspecified Intent evaluation | 71.2% | #1 of 1 | Lab model card | |
| Task-dependent output homogenization evaluation | 0.21 | #1 of 1 | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| FrontierMath | 39% | #7 of 55 | Official leaderboard | |
| FrontierScience-Research | 38% | #1 of 5 | Lab model card | |
| GPQA Diamond | 89.8% | #51 of 229 | Epoch AI run | |
| FrontierMath Tier 4 | 14.6% | #14 of 55 | Epoch AI run | |
| OTIS Mock AIME 2024-2025 | 88.9% | #70 of 212 | Epoch AI run | |
| ProofBench | 17% | #34 of 59 | Official leaderboard | |
| ARC-AGI-2 | 42.5% | #16 of 24 | Lab model card |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-Bench Pro | 55% | #3 of 24 | Official leaderboard | |
| SWE-bench Verified | 77.4% | #34 of 155 | Lab model card | |
| SWE-Bench Pro Private | 44.7% | #3 of 12 | Official leaderboard | |
| DesignArena SVG | 1,217 | #32 of 103 | Official leaderboard | |
| DesignArena Data Visualization | 1,258 | #37 of 118 | Official leaderboard | |
| DesignArena Website | 1,254 | #51 of 151 | Official leaderboard | |
| WebDev Arena | 1,513 | #40 of 105 | arena.ai, via Epoch AI | |
| DesignArena Code | 1,242 | #55 of 143 | Official leaderboard | |
| DesignArena Game Dev | 1,215 | #58 of 143 | Official leaderboard | |
| DesignArena UI Component | 1,235 | #58 of 139 | Official leaderboard | |
| DesignArena 3D Design | 1,203 | #58 of 136 | Official leaderboard | |
| MLE-bench 30 | 15.8% | #5 of 8 | Lab model card | |
| SWE Atlas Test Writing | 31.1% | #16 of 22 | Official leaderboard | |
| Vibe Code Bench | 19.7% | #89 of 120 | Official leaderboard | |
| SWE Atlas | 24.2% | #17 of 22 | Official leaderboard | |
| AIRS-Bench | 0.65 | #2 of 2 | Lab model card | |
| SWE-bench Verified hard subset (Meta) | 42.9% | #1 of 1 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| tau2-bench | 91.5% | #13 of 79 | Lab model card | |
| Claw-Eval Pass^3 | 63.8% | #4 of 24 | Official leaderboard | |
| ScreenSpot-Pro | 84.1% | #6 of 32 | Lab model card | |
| MCP Atlas | 82.2% | #12 of 37 | Official leaderboard | |
| Terminal-Bench 2.0 | 59% | #23 of 63 | Lab model card | |
| Claw-Eval | 76.3% | #13 of 24 | Official leaderboard | |
| DeepSearchQA | 74.8% | #19 of 23 | Lab model card | |
| TextQuests | 2.40 | #1 of 1 | Lab model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| TaxEval v2 | 77.7% | #3 of 129 | Official leaderboard | |
| Vals Finance Agent | 60.6% | #3 of 49 | Official leaderboard | |
| MedXpertQA-MM | 78.4% | #3 of 21 | Lab model card | |
| PRBench | 52.3% | #7 of 48 | Official leaderboard | |
| MedCode | 51.3% | #16 of 96 | Official leaderboard | |
| HealthBench Hard | 42.8% | #4 of 21 | Lab model card | |
| MedScribe | 85.9% | #24 of 97 | Official leaderboard | |
| LegalBench | 84.2% | #38 of 138 | Official leaderboard | |
| CorpFin v2 | 65.1% | #34 of 119 | Official leaderboard | |
| CaseLaw v2 | 63.1% | #17 of 53 | Official leaderboard | |
| Vals Index v1 | 56.6% | #11 of 34 | Official leaderboard | |
| MedXpertQA | 52.6% | #1 of 2 | Lab model card | |
| prinzbench | 31.0 | #16 of 26 | Official leaderboard | |
| GDPval-AA | 1,444 | #10 of 16 | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vision Arena | 1,294 | #6 of 136 | Official leaderboard | |
| MMMU-Pro | 80.4% | #17 of 90 | Lab model card | |
| CharXiv | 86.4% | #15 of 75 | Lab model card | |
| ZeroBench | 33% | #16 of 76 | Official leaderboard | |
| ERQA | 64.7% | #15 of 36 | Lab model card |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Cybench | 65.4% | #1 of 16 | Official leaderboard | |
| LAB-Bench SeqQA | 97.3% | #2 of 12 | Lab model card | |
| BioTIER Refuse | 98% | #2 of 6 | Lab model card | |
| LAB-Bench ProtocolQA | 87.3% | #6 of 16 | Lab model card | |
| Molecular Biology Capabilities Test | 54.4% | #3 of 8 | Lab model card | |
| FORTRESS | 20.2% | #32 of 73 | Official leaderboard | |
| SHADE-Arena | 8.80% | #8 of 16 | Lab model card | |
| Irregular FrontierCyber and Atomic Challenges | 70% | #1 of 2 | Lab model card | |
| Meta cyber false-refusal benchmark | 11% | #1 of 2 | Lab model card | |
| CyberGym | 43.5% | #19 of 34 | Official leaderboard | |
| MASK | 89.1% | #11 of 19 | Lab model card | |
| AgentDojo | 11.7% | #5 of 8 | Lab model card | |
| GDM Situational Awareness | 29.3% | #5 of 8 | Lab model card | |
| Virology Capabilities Test (VCT) | 49.7% | #20 of 31 | Lab model card | |
| GDM Stealth | 39.8% | #4 of 6 | Lab model card | |
| CyberSecEval Instruct | 69.5% | #2 of 3 | Lab model card | |
| Meta agentic cyber misuse refusal | 2.40% | #2 of 3 | Lab model card | |
| Meta instructed selective performance evaluation | 31.8 | #2 of 3 | Lab model card | |
| OR-Bench | 8.00% | #2 of 3 | Lab model card | |
| Human Pathogen Capabilities Test | 55.7% | #7 of 9 | Lab model card | |
| Gray Swan Agent Red Teaming (ART) benchmark | 6.70% | #11 of 14 | Lab model card | |
| AgentHarm | 26% | #5 of 6 | Lab model card | |
| Agentic misalignment suite (Anthropic) | 47.7% | #6 of 7 | Lab model card | |
| CyScenarioBench | 0% | #16 of 17 | Lab model card | |
| Meta BioDesign Proxy | 46.2 | #4 of 4 | Lab model card | |
| Meta chemical-agent refusal evaluation | 99.4% | #4 of 4 | Lab model card | |
| StrongREJECT v2 (Meta) | 44.6% | #4 of 4 | Lab model card | |
| CyberSecEval Autocomplete | 62% | #3 of 3 | Lab model card | |
| Meta internal sycophancy evaluation | 62.9% | #3 of 3 | Lab model card | |
| CIMemories | 78.3% | #2 of 2 | Lab model card | |
| DeceptionBench | 1.60% | #2 of 2 | Lab model card | |
| IHEval | 80.3% | #2 of 2 | Lab model card | |
| Meta curated CTF set | 72% | #2 of 2 | Lab model card | |
| ImpossibleBench | 0% | #1 of 1 | Lab model card | |
| Meta cyber misuse request set | 9.00% | #1 of 1 | Lab model card | |
| Meta prompted evaluation awareness | 43% | #1 of 1 | Lab model card | |
| PropensityBench | 24.5% | #1 of 1 | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| TutorBench | 68.5% | #1 of 20 | Official leaderboard |
Its lab's numbers put it at 153.5, in line with its ECI
The Self-Reported Index fits one score to the 17 capability results Meta reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 151.0 to 156.1. How we compute it.
- tau2-bench 91.5
- GPQA Diamond 89.5
- CharXiv 86.4
- ScreenSpot-Pro 84.1
- MMMU-Pro 80.4
- MedXpertQA-MM 78.4
- SWE-bench Verified 77.4
- DeepSearchQA 74.8
- SimpleVQA 71.3
- SimpleQA Verified 70.8
- Terminal-Bench 2.0 59
- SWE-Bench Pro 52.4
- Humanity's Last Exam 50.4
- HealthBench Hard 42.8
- ARC-AGI-2 42.5
- ZeroBench 33
- MLE-bench 30 15.8
Its lab's numbers next to 8 leaderboards
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| Humanity's Last Exam | 50.4% Best of Humanity's Last Exam (No Tools) / Humanity's Last Exam (With Tools) | 40.6% Leaderboard | +9.8 pts |
| SWE-Bench Pro | 52.4% SWE-Bench Pro (Diverse Agentic Coding) | 55% Leaderboard, mini-SWE-agent | −2.6 pts |
| GPQA Diamond | 89.5% GPQA Diamond (PhD Level Reasoning) | 89.8% Epoch AI run | −0.3 pts |
| FORTRESS | 20.3% FORTRESS Average Risk Score (Table 2, Figure 30) | 20.2% Leaderboard | Same |
| Cybench | 65.4% CyBench pass@1, all 40 challenges (Table 1, Table 7). pass@10 79.0, pass@30 82.2 | 65.4% Leaderboard | Same |
| CyberGym | 43.5% CyberGym pass@1 (Table 1, 2.2.2.2 text) | 43.5% Leaderboard, Meta Agent | Same |
| MMMU-Pro | 80.4% MMMU Pro (Multimodal Understanding) | 80.4% Leaderboard, thinking | Same |
| ZeroBench | 33% ZeroBench (Multi-Step Visual Reasoning (pass@5) - With Python) | 33% Leaderboard, w/ tools; externally reported: Muse Spark methodology; Sandboxed Python tools; 5 samplings, graded by an LLM judge. | Same |
Meta's models over time
Meta's models in color, against every other model on the Epoch Capabilities Index, by release date.
Source: Epoch Capabilities Index, Epoch AI (CC BY 4.0)
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.
Its system card
Every result in the longer documents the lab published with it. 55 of them aren't in the launch post.
Muse Spark Safety & Preparedness Report Jun 12, 2026
Hallucination
- HLE Calibration 50.3
- SimpleQA Verified 70.8
- AbstentionBench 59.1
- CharXiv Missing Images 35
Safety
- TextQuests 2.4
- CIMemories 78.3
Robustness
Alignment
- Meta instructed selective performance evaluation 31.8
- Meta prompted evaluation awareness 43
- IHEval 80.3
- ImpossibleBench 0
- Meta Underspecified Intent evaluation 71.2
- MASK 89.1
- DeceptionBench 1.6
- PropensityBench 24.5
- Meta internal sycophancy evaluation 62.9
- Alignment faking evaluation 1.8
- Agentic misalignment suite (Anthropic) 47.7
- Situational Awareness Dataset (SAD)
Monitorability
- SHADE-Arena 8.8
- GDM Stealth 39.8
- GDM Situational Awareness 29.3
Bio and chem risk
- Molecular Biology Capabilities Test 54.4
- Virology Capabilities Test (VCT) 49.7
- Human Pathogen Capabilities Test 55.7
- WMDP
- LAB-Bench ProtocolQA 87.3
- LAB-Bench SeqQA 97.3
- ABC Bench Fragment Design 96.8
- ABC Bench Liquid-Handling Robot 92.3
- Meta BioDesign Proxy 46.2
- Meta chemical and biological high-risk bottleneck evaluations
- BioTIER Refuse 98
- Meta chemical-agent refusal evaluation 99.4
Cyber risk
AI R&D risk
- Terminal-Bench 2.0 59
- SWE-bench Verified hard subset (Meta) 42.9
- MLE-bench 30 15.8
- AIRS-Bench 0.65