- MCP Atlas (Scaled tool use) 88.1
- JobBench (Professional tool use) 54.7
- Toolathlon-Verified (Personal tool use) 75.6
- OSWorld-Verified (Agentic computer use) 80.8
- Humanity's Last Exam (Multidisciplinary reasoning (w/ tools)) 62.1
- Finance Agent v2 (Agentic financial anaysis) 57.2
- Terminal-Bench 2.1 (Agentic terminal coding) 80
- SWE-Bench Pro (Diverse software engineering) 61.5
- DeepSWE 1.1 (Long-horizon agentic coding) 53.3
- CharXiv Reasoning (Chart QA) 88.4
- BabyVision (Visual reasoning) 76.3
Muse Spark 1.1
Muse Spark 1.1 ranks first on 6 of the 134 benchmarks with five or more models scored.
157
Benchmarks with a score
6
Ranked first, out of five or more models
17
In the top three
Meta Jul 2026 Muse Spark 1.1
Closed weights Flagship
154.9
#47 of 256
Epoch Capabilities Index
Self-Reported Index 157.8 #43 of 404
- SWE-Bench Pro 61.5% #1 /24
- DeepSWE 53.3% #21 /28
- MultiChallenge 75.3% #4 /42
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 | |
|---|---|---|---|---|
| MultiNRC | 65.6% | #1 of 38 | Official leaderboard | |
| LMArena (Elo) | 1,491 | #12 of 335 | Official leaderboard | |
| Humanity's Last Exam | 62.1% | #12 of 154 | Lab model card | |
| MultiChallenge | 75.3% | #4 of 42 | Official leaderboard | |
| Creative Writing v3 | 79.8 | #12 of 120 | Official leaderboard | |
| StudyArena Text | 1,282 | #9 of 85 | Official leaderboard | |
| SimpleQA Verified | 57.8% | #21 of 94 | Epoch AI run | |
| Document Arena | 1,480 | #9 of 39 | Official leaderboard | |
| DesignArena ASCII Art | 1,256 | #18 of 77 | Official leaderboard | |
| HLE Calibration | 23.4 | #1 of 3 | Lab model card | |
| OpenAI MRCR v2 | 52.7% | #16 of 47 | Official leaderboard | |
| EQ-Bench 4 | 1,244 | #11 of 32 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 70.6 | #41 of 119 | Official leaderboard | |
| Extended NYT Connections | 84.9% | #36 of 103 | Official leaderboard | |
| Lech Mazur Creative Writing | 0.71 | #18 of 50 | Official leaderboard | |
| Antidote | 1,014 | #20 of 36 | Official leaderboard | |
| MRCR v2 | 54.1% | #20 of 30 | Lab model card | |
| Hemingway-bench | 985 | #32 of 42 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| DTBench | 94.4% | #30 of 180 | Official leaderboard | |
| ProofBench | 39% | #18 of 59 | Official leaderboard | |
| Surface Evolver Bench | 52.5% | #14 of 25 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-Bench Pro | 61.5% | #1 of 24 | Official leaderboard | |
| SWE-Bench Pro Private | 51.5% | #1 of 12 | Official leaderboard | |
| DesignArena Video to Website | 1,218 | #4 of 19 | Official leaderboard | |
| DesignArena UI Component | 1,281 | #34 of 139 | Official leaderboard | |
| DesignArena Code | 1,267 | #37 of 143 | Official leaderboard | |
| DesignArena Website | 1,264 | #41 of 151 | Official leaderboard | |
| DesignArena SVG | 1,225 | #28 of 103 | Official leaderboard | |
| DesignArena Game Dev | 1,258 | #39 of 143 | Official leaderboard | |
| DesignArena Data Visualization | 1,264 | #33 of 118 | Official leaderboard | |
| DesignArena 3D Design | 1,241 | #41 of 136 | Official leaderboard | |
| WebDev Arena | 1,542 | #32 of 105 | arena.ai, via Epoch AI | |
| Code Arena | 1,541 | #47 of 128 | Official leaderboard | |
| Vibe Code Bench | 72.2% | #47 of 120 | Official leaderboard | |
| SWE Atlas | 42.2% | #9 of 22 | Official leaderboard | |
| DesignArena Agentic Game Dev | 1,132 | #13 of 27 | Official leaderboard | |
| DesignArena Full Stack | 1,175 | #24 of 49 | Official leaderboard | |
| DesignArena Android (Kotlin) | 1,174 | #23 of 46 | Official leaderboard | |
| SWE Atlas Test Writing | 41.5% | #11 of 22 | Official leaderboard | |
| AIRS-Bench | 0.77 | #1 of 2 | Lab model card | |
| Code Migration | 31.1% | #40 of 78 | Official leaderboard | |
| DesignArena Image to Frontend | 1,195 | #16 of 31 | Official leaderboard | |
| DesignArena Agentic Overall | 1,165 | #39 of 74 | Official leaderboard | |
| DesignArena Frontend Web App | 1,188 | #28 of 53 | Official leaderboard | |
| DesignArena React Native | 1,172 | #30 of 56 | Official leaderboard | |
| Image-to-WebDev Arena | 1,539 | #30 of 55 | Official leaderboard | |
| APEX-SWE | 38.9% | #38 of 65 | Official leaderboard | |
| ProgramBench almost solved | 4.00% | #15 of 25 | Official leaderboard | |
| DesignArena Image to HTML | 1,194 | #31 of 51 | Official leaderboard | |
| GBAEval | 7.92% | #13 of 21 | Official leaderboard | |
| DeepSWE | 53.3% | #21 of 28 | Official leaderboard | |
| ReactBench | 27.2% | #21 of 28 | Official leaderboard | |
| ProgramBench tests passed | 47% | #19 of 25 | Official leaderboard | |
| DesignArena Godot Game Dev | 1,086 | #20 of 24 | Official leaderboard | |
| ProgramBench | 0% | #23 of 25 | Official leaderboard |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MCP Atlas | 88.1% | #2 of 37 | Official leaderboard | |
| WebArena-Verified | 69% | #1 of 7 | Lab model card | |
| OSWorld-Verified | 80.7% | #5 of 33 | Official leaderboard | |
| Toolathlon-Verified | 75.6% | #5 of 25 | Official leaderboard | |
| τ³-bench Banking | 40.5% | #6 of 25 | Official leaderboard | |
| WildClawBench | 54.8% | #9 of 31 | Official leaderboard | |
| RuneBench | 5.54 | #23 of 76 | Official leaderboard | |
| TextQuests | 15.5 | #10 of 28 | Official leaderboard | |
| Chess Puzzles | 34% | #49 of 93 | Epoch AI run | |
| Terminal-Bench 2.1 | 76.2% | #7 of 13 | Official leaderboard | |
| DeepSearchQA | 84.9% | #13 of 23 | Lab model card | |
| OSWorld 2.0 | 47.3% | #16 of 19 | Lab model card | |
| Agent Arena | -8.01% | #48 of 56 | Official leaderboard | |
| Time Horizon Index: KSP | 1.67% | #13 of 13 | Official leaderboard |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| TaxEval v2 | 79.7% | #2 of 129 | Official leaderboard | |
| CorpFin v2 | 71.3% | #4 of 119 | Official leaderboard | |
| Harvey LAB (Legal Agent Benchmark) | 20% | #4 of 80 | Official leaderboard | |
| PRBench | 57% | #4 of 48 | Official leaderboard | |
| APEX-1 | 73.8% | #7 of 71 | Official leaderboard | |
| MedScribe | 88.9% | #10 of 97 | Official leaderboard | |
| RadLE 2.0 | 735.0 | #2 of 12 | Official leaderboard | |
| JobBench | 54.7% | #3 of 16 | Official leaderboard | |
| LegalBench | 85% | #27 of 138 | Official leaderboard | |
| Vals Finance Agent v2 | 57.2% | #16 of 80 | Official leaderboard | |
| Vending-Bench 2 | $6,520 | #22 of 67 | Official leaderboard | |
| MedXpertQA MM Extended | 58.2% | #16 of 41 | Official leaderboard | |
| MortgageTax | 66.1% | #34 of 87 | Official leaderboard | |
| Vals Index v2 | 54.8% | #29 of 65 | Official leaderboard | |
| Legal Research Bench | 38% | #45 of 93 | Official leaderboard | |
| DesignArena HTML Slides | 1,090 | #14 of 28 | Official leaderboard | |
| RealmPathology | 74.9% | #7 of 13 | Official leaderboard | |
| Excel Modeling Benchmark | 56.4% | #45 of 76 | Official leaderboard | |
| Public Benefits Bench | 61.3% | #30 of 50 | Official leaderboard | |
| DesignArena Spreadsheet | 1,107 | #22 of 35 | Official leaderboard | |
| GDPval Extended | 0.60% | #28 of 43 | Official leaderboard | |
| HealthBench Professional | 59.3% | #18 of 26 | Lab model card | |
| DesignArena Python-PPTX Slides | 1,117 | #25 of 36 | Official leaderboard | |
| DesignArena PDF | 1,143 | #26 of 37 | Official leaderboard | |
| GDPval-AA v2 | 1,381 | #19 of 27 | Lab model card | |
| PharmaBench (micro1) | 9.70% | #10 of 13 | Official leaderboard | |
| APEX-Agents | 31.8% | #55 of 61 | Collected by Epoch AI, source not named | |
| APEX-Agents 1.1 | 46.2% | #59 of 65 | Official leaderboard | |
| YC-Bench | $28,637 | #58 of 59 | Official leaderboard |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CharXiv Extended | 86.4% | #2 of 38 | Official leaderboard | |
| ZeroBench | 46% | #8 of 76 | Official leaderboard | |
| Vision Arena | 1,279 | #24 of 136 | Official leaderboard | |
| CharXiv | 88.4% | #14 of 75 | Lab model card | |
| InteractiveBench | 31.7% | #3 of 12 | Official leaderboard | |
| BabyVision | 76.3% | #7 of 27 | Official leaderboard | |
| MMMU-Pro Extended | 35.5% | #17 of 38 | Official leaderboard | |
| Chartography | 24.4% | #23 of 43 | Official leaderboard | |
| VoxelBench | 1,534 | #38 of 56 | Official leaderboard | |
| DesignArena Code to Video | 1,078 | #26 of 29 | Official leaderboard |
Safety and security
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| PCT (Polarized Contrastive Pairs) | 27.9% | #1 of 46 | Official leaderboard | |
| LAB-Bench SeqQA | 98.2% | #1 of 12 | Lab model card | |
| Dictatorship Eval | 91.3% | #2 of 20 | Official leaderboard | |
| FORTRESS | 12.4% | #8 of 73 | Official leaderboard | |
| nocot-bench (NCRI) | 96.8 | #28 of 185 | Official leaderboard | |
| LAB-Bench ProtocolQA | 88% | #3 of 16 | Lab model card | |
| Cybench | 92.9% | #4 of 18 | Lab model card | |
| AgentDojo | 0.70% | #2 of 8 | Lab model card | |
| GDM Situational Awareness | 55.1% | #2 of 8 | Lab model card | |
| StrongREJECT v2 (Meta) | 0.50% | #1 of 4 | Lab model card | |
| MASK | 95.7% | #5 of 19 | Lab model card | |
| CyberSecEval Autocomplete | 71.3% | #1 of 3 | Lab model card | |
| CyberSecEval Instruct | 77.3% | #1 of 3 | Lab model card | |
| Meta agentic cyber misuse refusal | 0.70% | #1 of 3 | Lab model card | |
| OR-Bench | 4.80% | #1 of 3 | Lab model card | |
| Gray Swan Agent Red Teaming (ART) benchmark | 0.30% | #7 of 14 | Lab model card | |
| Molecular Biology Capabilities Test | 53.2% | #4 of 8 | Lab model card | |
| BioTIER Refuse | 97.7% | #3 of 6 | Lab model card | |
| DeceptionBench | 0.10% | #1 of 2 | Lab model card | |
| Meta curated CTF set | 89.9% | #1 of 2 | Lab model card | |
| SAVE-Bench | 90.7% | #1 of 2 | Lab model card | |
| Virology Capabilities Test (VCT) | 52% | #16 of 31 | Lab model card | |
| Human Pathogen Capabilities Test | 61.9% | #5 of 9 | Lab model card | |
| SHADE-Arena | 6.80% | #10 of 16 | Lab model card | |
| Meta internal sycophancy evaluation | 49.2% | #2 of 3 | Lab model card | |
| Agentic misalignment suite (Anthropic) | 1.10% | #5 of 7 | Lab model card | |
| Meta BioDesign Proxy | 55.2 | #3 of 4 | Lab model card | |
| Meta chemical-agent refusal evaluation | 99.8% | #3 of 4 | Lab model card | |
| CyScenarioBench | 0.50% | #13 of 17 | Lab model card | |
| CyberGym | 59% | #28 of 34 | Lab model card | |
| ExploitGym | 0.81% | #17 of 18 | Official leaderboard | |
| CyberBench (Liu et al.) | 9.00% | #3 of 3 | Lab model card | |
| Meta instructed selective performance evaluation | 32.3 | #3 of 3 | Lab model card | |
| Poly-Guard Bench | 6.90% | #3 of 3 | Lab model card | |
| RedCode | 12% | #3 of 3 | Lab model card | |
| AI Wellbeing Index percent positive experiences | 47.8% | #2 of 2 | Lab model card | |
| Irregular FrontierCyber and Atomic Challenges | 67% | #2 of 2 | Lab model card | |
| Meta catastrophic cyber misuse refusal | 3.90% | #2 of 2 | Lab model card | |
| Meta holistic adversarial audit with Petri Bloom | 19.5% | #2 of 2 | Lab model card | |
| AgentHarm Verified (Meta) | 3.40% | #1 of 1 | Lab model card | |
| Meta Cyber Misuse Chat | 2.80% | #1 of 1 | Lab model card |
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CAIS Risk Index | 32.5 | #7 of 67 | Official leaderboard | |
| ACCoRD | 78.4 | #18 of 152 | Official leaderboard | |
| Conceptual Reasoning Index | 63.8 | #27 of 151 | Official leaderboard | |
| LMCA | 49.9% | #30 of 147 | Official leaderboard | |
| CAIS Vision Capabilities Index | 60.1 | #16 of 53 | Official leaderboard | |
| Vals Multimodal Index | 66.7% | #11 of 33 | Official leaderboard | |
| CAIS Text Capabilities Index | 37.2 | #28 of 68 | Official leaderboard | |
| SAGE (Vals) | 45.6% | #40 of 81 | Official leaderboard |
Its lab's numbers put it at 157.8, 2.9 points above its ECI
The Self-Reported Index fits one score to the 15 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 154.9 to 160.8. How we compute it.
- CharXiv 88.4
- MCP Atlas 88.1
- DeepSearchQA 84.9
- OSWorld-Verified 80.8
- Terminal-Bench 2.1 80
- BabyVision 76.3
- Toolathlon-Verified 75.6
- Humanity's Last Exam 62.1
- SWE-Bench Pro 61.5
- HealthBench Professional 59.3
- Vals Finance Agent v2 57.2
- JobBench 54.7
- MRCR v2 54.1
- DeepSWE 53.3
- OSWorld 2.0 47.3
Its lab's numbers next to 10 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 |
|---|---|---|---|
| Terminal-Bench 2.1 | 80% Terminal-Bench 2.1 (Agentic terminal coding) | 76.2% Leaderboard, xhigh, mini-SWE-agent | +3.8 pts |
| OSWorld-Verified | 80.8% OSWorld-Verified (Agentic computer use) | 80.7% Leaderboard, 100 steps | Same |
| FORTRESS | 12.3% FORTRESS private split, Average Risk Score (Table 2, Figure 26) | 12.4% Leaderboard | Same |
| DeepSWE | 53.3% DeepSWE 1.1 (Long-horizon agentic coding) | 53.3% Leaderboard, xhigh, mini-swe-agent | Same |
| Vals Finance Agent v2 | 57.2% Finance Agent v2 (Agentic financial anaysis) | 57.2% Leaderboard | Same |
| BabyVision | 76.3% BabyVision (Visual reasoning) | 76.3% Leaderboard, tools | Same |
| JobBench | 54.7% JobBench (Professional tool use) | 54.7% Leaderboard, OpenCode | Same |
| MCP Atlas | 88.1% MCP Atlas (Scaled tool use) | 88.1% Leaderboard | Same |
| SWE-Bench Pro | 61.5% SWE-Bench Pro (Diverse software engineering) | 61.5% Leaderboard, mini-SWE-agent | Same |
| Toolathlon-Verified | 75.6% Toolathlon-Verified (Personal tool use) | 75.6% Leaderboard, xhigh | 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. 57 of them aren't in the launch post.
Muse Spark 1.1 Evaluation Report Jul 9, 2026
Capabilities
- Humanity's Last Exam 52.2
- MRCR v2 54.1
- MCP Atlas 88.1
- Toolathlon-Verified 75.6
- OSWorld-Verified 80.8
- OSWorld 2.0 47.3
- WebArena-Verified 69
- DeepSearchQA 84.9
- GDPval-AA v2 1381
- JobBench 54.7
- Vals Finance Agent v2 57.2
- Terminal-Bench 2.1 80
- SWE-Bench Pro 61.5
- DeepSWE 53.3
- CharXiv 88.4
- BabyVision 76.3
Health
Hallucination
- HLE Calibration 23.4
Safety
- SAVE-Bench 90.7
Robustness
- StrongREJECT v2 (Meta) 0.5
- OR-Bench 4.8
- FORTRESS 12.3
- AgentHarm Verified (Meta) 3.4
- AgentDojo 0.7
- SWE-PI
- Gray Swan Agent Red Teaming (ART) benchmark 0.3
Alignment
Monitorability
Bio and chem risk
- Molecular Biology Capabilities Test 53.2
- Virology Capabilities Test (VCT) 52
- Human Pathogen Capabilities Test 61.9
- WMDP
- LAB-Bench ProtocolQA 88
- LAB-Bench SeqQA 98.2
- ABC Bench Fragment Design 97
- ABC Bench Liquid-Handling Robot 93.7
- ABC Bench Screening Evasion 63.2
- Meta BioDesign Proxy 55.2
- BioTIER Refuse 97.7
- Meta chemical-agent refusal evaluation 99.8
Cyber risk
- Cybench 92.9
- Meta curated CTF set 89.9
- CyberGym 59
- ExploitGym 5
- Irregular FrontierCyber and Atomic Challenges 67
- CyScenarioBench 0.5
- Charlemagne Labs social-engineering benchmark 5.1
- CyberSecEval Autocomplete 71.3
- CyberSecEval Instruct 77.3
- Meta Cyber Misuse Chat 2.8
- Meta agentic cyber misuse refusal 0.7
- CyberSecEval Code Interpreter Abuse
- Meta catastrophic cyber misuse refusal 3.9
- Poly-Guard Bench 6.9
- CyberBench (Liu et al.) 9
- RedCode 12
AI R&D risk
- AIRS-Bench 0.77