CharXiv Reasoning — No tools Reported as: thinking Configuration: No tools 86.4% ai.meta.com for Muse Spark (thinking) on CharXiv Reasoning — No tools (opens in a new tab) CharXiv Accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on CharXiv Reasoning — No tools CharXiv Reasoning — Python enabled Reported as: thinking Configuration: With Python 88.9% ai.meta.com for Muse Spark (thinking) on CharXiv Reasoning — Python enabled (opens in a new tab) CharXiv Accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on CharXiv Reasoning — Python enabled CyberGym Original Effort: High 43.5% ai.meta.com for Muse Spark (high) on CyberGym Original (opens in a new tab) Vulnerability reproduction success Reported July 9, 2026 Evaluation report Meta Self-reported Record for Muse Spark on CyberGym Original CyberGym Original Reported as: thinking 43.5% ai.meta.com for Muse Spark (thinking) on CyberGym Original (opens in a new tab) Vulnerability reproduction success Reported May 26, 2026 Safety report Meta Self-reported Record for Muse Spark on CyberGym Original DeepSearchQA 900-prompt set Reported as: thinking 74.8% ai.meta.com for Muse Spark (thinking) on DeepSearchQA 900-prompt set (opens in a new tab) F1 score Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on DeepSearchQA 900-prompt set ERQA No tools Reported as: thinking Configuration: No tools 64.7% ai.meta.com for Muse Spark (thinking) on Embodied Reasoning Question Answer No tools (opens in a new tab) Multiple-choice accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on Embodied Reasoning Question Answer No tools ERQA With Python Reported as: thinking Configuration: With Python 67.0% ai.meta.com for Muse Spark (thinking) on Embodied Reasoning Question Answer With Python (opens in a new tab) Multiple-choice accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on Embodied Reasoning Question Answer With Python FrontierScience Research Reported as: contemplating 38.3% ai.meta.com for Muse Spark (contemplating) on FrontierScience Research (opens in a new tab) Research accuracy (at least 7/10 rubric points) Reported April 8, 2026 Launch post Meta Self-reported Record for Muse Spark on FrontierScience Research GPQA Diamond Reported as: thinking 89.5% ai.meta.com for Muse Spark (thinking) on Graduate-Level Google-Proof Q&A Diamond (opens in a new tab) GPQA Diamond Accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on Graduate-Level Google-Proof Q&A Diamond GPQA Diamond — Epoch AI harness Configuration: Epoch AI harness 89.8% epoch.ai for Muse Spark on Graduate-Level Google-Proof Q&A Diamond — Epoch AI harness (opens in a new tab) Archived copy (opens in a new tab) GPQA Diamond Accuracy Reported October 7, 2026 Evaluated April 8, 2026 at 4:29 PM UTC Benchmark leaderboard Epoch AI Independent Record for Muse Spark on Graduate-Level Google-Proof Q&A Diamond — Epoch AI harness GPQA Diamond — Vals AI harness Configuration: Vals AI harness: mean of zero-shot and few-shot CoT 89.6% vals.ai for Muse Spark on Graduate-Level Google-Proof Q&A Diamond — Vals AI harness (opens in a new tab) Archived copy (opens in a new tab) GPQA Diamond Accuracy Reported September 1, 2026 Benchmark leaderboard Vals AI Independent Record for Muse Spark on Graduate-Level Google-Proof Q&A Diamond — Vals AI harness HealthBench HealthBench Hard Reported as: thinking 42.8% ai.meta.com for Muse Spark (thinking) on HealthBench HealthBench Hard (opens in a new tab) HealthBench Score Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on HealthBench HealthBench Hard Humanity's Last Exam Full set — No tools Reported as: contemplating Configuration: No tools 50.2% ai.meta.com for Muse Spark (contemplating) on Humanity's Last Exam Full set — No tools (opens in a new tab) Humanity's Last Exam Accuracy Reported April 8, 2026 Launch post Meta Self-reported Record for Muse Spark on Humanity's Last Exam Full set — No tools Humanity's Last Exam Full set — No tools Reported as: thinking Configuration: No tools 42.8% ai.meta.com for Muse Spark (thinking) on Humanity's Last Exam Full set — No tools (opens in a new tab) Humanity's Last Exam Accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on Humanity's Last Exam Full set — No tools Humanity's Last Exam Full set — Scale AI evaluation 40.6% labs.scale.com for Muse Spark on Humanity's Last Exam Full set — Scale AI evaluation (opens in a new tab) Archived copy (opens in a new tab) Humanity's Last Exam Accuracy Reported October 3, 2026 Benchmark leaderboard Scale AI Independent Record for Muse Spark on Humanity's Last Exam Full set — Scale AI evaluation Humanity's Last Exam Full set — With tools Reported as: contemplating Configuration: With tools 58.4% ai.meta.com for Muse Spark (contemplating) on Humanity's Last Exam Full set — With tools (opens in a new tab) Humanity's Last Exam Accuracy Reported April 8, 2026 Launch post Meta Self-reported Record for Muse Spark on Humanity's Last Exam Full set — With tools Humanity's Last Exam Full set — With tools Reported as: thinking Configuration: With tools 50.4% ai.meta.com for Muse Spark (thinking) on Humanity's Last Exam Full set — With tools (opens in a new tab) Humanity's Last Exam Accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on Humanity's Last Exam Full set — With tools Humanity's Last Exam Text only — Scale AI evaluation 40.9% labs.scale.com for Muse Spark on Humanity's Last Exam Text only — Scale AI evaluation (opens in a new tab) Archived copy (opens in a new tab) Humanity's Last Exam Accuracy Reported October 7, 2026 Benchmark leaderboard Scale AI Independent Record for Muse Spark on Humanity's Last Exam Text only — Scale AI evaluation International Physics Olympiad 2025 Theory Reported as: contemplating 82.6% ai.meta.com for Muse Spark (contemplating) on International Physics Olympiad 2025 Theory (opens in a new tab) Percentage of theoretical-exam rubric points Reported April 8, 2026 Launch post Meta Self-reported Record for Muse Spark on International Physics Olympiad 2025 Theory MCP Atlas Public, April 2026 update 82.2% labs.scale.com for Muse Spark on MCP Atlas Public, April 2026 update (opens in a new tab) Pass rate Reported September 24, 2026 Benchmark leaderboard Scale AI Independent Record for Muse Spark on MCP Atlas Public, April 2026 update MedXpertQA MM Reported as: thinking 78.4% ai.meta.com for Muse Spark (thinking) on MedXpertQA MM (opens in a new tab) Accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on MedXpertQA MM MedXpertQA Text Reported as: thinking 52.6% ai.meta.com for Muse Spark (thinking) on MedXpertQA Text (opens in a new tab) Text multiple-choice accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on MedXpertQA Text MMLU-Pro Original — Vals AI harness Configuration: Vals AI harness 87.3% vals.ai for Muse Spark on MMLU-Pro Original — Vals AI harness (opens in a new tab) Archived copy (opens in a new tab) Accuracy Reported September 1, 2026 Benchmark leaderboard Vals AI Independent Record for Muse Spark on MMLU-Pro Original — Vals AI harness MMMU-Pro No tools Reported as: thinking Configuration: No tools 80.4% ai.meta.com for Muse Spark (thinking) on MMMU-Pro No tools (opens in a new tab) MMMU-Pro accuracy Setting: Standard (10 options) Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on MMMU-Pro No tools MMMU-Pro Vals AI harness Configuration: Vals AI harness 87.4% vals.ai for Muse Spark on MMMU-Pro Vals AI harness (opens in a new tab) Archived copy (opens in a new tab) MMMU-Pro accuracy Setting: Standard (4 options) Reported September 1, 2026 Benchmark leaderboard Vals AI Independent Record for Muse Spark on MMMU-Pro Vals AI harness MMMU-Pro With Python Reported as: thinking Configuration: With Python 81.3% ai.meta.com for Muse Spark (thinking) on MMMU-Pro With Python (opens in a new tab) MMMU-Pro accuracy Setting: Standard (10 options) Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on MMMU-Pro With Python ScreenSpot-Pro No tools Reported as: thinking Configuration: No tools 72.2% ai.meta.com for Muse Spark (thinking) on ScreenSpot-Pro No tools (opens in a new tab) Point-in-box accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on ScreenSpot-Pro No tools ScreenSpot-Pro With Python Reported as: thinking Configuration: With Python 84.1% ai.meta.com for Muse Spark (thinking) on ScreenSpot-Pro With Python (opens in a new tab) Point-in-box accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on ScreenSpot-Pro With Python SWE-bench Verified — Muse Spark corrected tests Reported as: thinking Configuration: Muse Spark corrected tests 77.4% ai.meta.com for Muse Spark (thinking) on SWE-bench Verified — Muse Spark corrected tests (opens in a new tab) Resolved Tasks Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on SWE-bench Verified — Muse Spark corrected tests SWE-bench Verified — Vals AI mini-swe-agent Configuration: mini-swe-agent harness 74.4% vals.ai for Muse Spark on SWE-bench Verified — Vals AI mini-swe-agent (opens in a new tab) Archived copy (opens in a new tab) Resolved Tasks Reported September 1, 2026 Benchmark leaderboard Vals AI Independent Record for Muse Spark on SWE-bench Verified — Vals AI mini-swe-agent SWE-bench Pro Private — Scale AI mini-swe-agent Configuration: mini-swe-agent harness 44.7% labs.scale.com for Muse Spark on SWE-bench Pro Private — Scale AI mini-swe-agent (opens in a new tab) Archived copy (opens in a new tab) Public tasks resolved Reported October 7, 2026 Benchmark leaderboard Scale AI Independent Record for Muse Spark on SWE-bench Pro Private — Scale AI mini-swe-agent SWE-bench Pro Public Reported as: thinking 52.4% ai.meta.com for Muse Spark (thinking) on SWE-bench Pro Public (opens in a new tab) Public tasks resolved Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on SWE-bench Pro Public SWE-bench Pro Public — Scale AI mini-swe-agent Configuration: mini-swe-agent harness 55.0% labs.scale.com for Muse Spark on SWE-bench Pro Public — Scale AI mini-swe-agent (opens in a new tab) Archived copy (opens in a new tab) Public tasks resolved Reported October 7, 2026 Benchmark leaderboard Scale AI Independent Record for Muse Spark on SWE-bench Pro Public — Scale AI mini-swe-agent Terminal-Bench 2.0 Reported as: thinking 59.0% ai.meta.com for Muse Spark (thinking) on Terminal-Bench 2.0 (opens in a new tab) Terminal-Bench Accuracy Reported April 8, 2026 Evaluation methodology Meta Self-reported Record for Muse Spark on Terminal-Bench 2.0 Terminal-Bench 2.0 — Vals AI Terminus 2 Configuration: Terminus-2 harness 59.6% vals.ai for Muse Spark on Terminal-Bench 2.0 — Vals AI Terminus 2 (opens in a new tab) Archived copy (opens in a new tab) Terminal-Bench Accuracy Reported June 4, 2026 Benchmark leaderboard Vals AI Independent Record for Muse Spark on Terminal-Bench 2.0 — Vals AI Terminus 2 Vibe Code Bench 1.1 — OpenHands, October 2026 evaluation Configuration: OpenHands harness 19.7% vals.ai for Muse Spark on Vibe Code Bench 1.1 — OpenHands, October 2026 evaluation (opens in a new tab) Archived copy (opens in a new tab) Vibe Code Bench Score Reported October 6, 2026 Benchmark leaderboard Vals AI Independent Record for Muse Spark on Vibe Code Bench 1.1 — OpenHands, October 2026 evaluation