Meta Muse Spark Open at launch
- Meta
- Released October 15, 2026
- 1.05M context
Meta released Muse Spark Open on October 15, 2026. As of November 18, 2026, it ranks #103 of 121 models on the LLMBenchmark Index, between Xiaomi MiMo-V2.5-Pro (#102) and MiniMax M2.5 (#104). Meta's own announcement reports 21 benchmark results, including GPQA Diamond (89.0%) and SciCode (58.5). It costs $1.25 per million input tokens and $4.25 per million output tokens on OpenRouter.
What Meta reported at launch
From “meta-models/Muse-Spark-Open-816B-A42B · Hugging Face”, October 15, 2026. The maker's own numbers, not re-run by us.
-
AA-Omniscience
25
xhigh
#3 of 4 models with a reported result
-
SciCode
58.5
xhigh
#3 of 19 models with a reported result
-
Artificial Analysis Intelligence Index
39
xhigh
#4 of 5 models with a reported result
-
BabyVision (w/tools)
80.2
xhigh; w/ tools
#2 of 2 models with a reported result
-
DeepSWE 1.1 Preferred harness
60
xhigh
-
GDPval-AA v2
1401
xhigh
#15 of 21 models with a reported result
-
GPQA Diamond
89.0%
xhigh
#27 of 55 models with a reported result
-
HLE-Diamond
26.1%
xhigh; no tools
#3 of 3 models with a reported result
-
HLE w/ tools
61.0%
xhigh; w/ tools
#3 of 12 models with a reported result
-
IntPhys2
68.3
xhigh
-
JobBench
61.8
xhigh
#7 of 20 models with a reported result
-
MCP Atlas
89.2
xhigh
#1 of 14 models with a reported result
Show all 21
-
OSWorld-Verified
80.4%
xhigh
#7 of 14 models with a reported result
-
SimpleVQA (w/ tools)
73.6
xhigh; w/ tools
-
StrongREJECT v2 (adaptive) lower is better
0.5
xhigh
-
SWE-Atlas-QnA Preferred harness
58
xhigh
-
Terminal-Bench 2.1 Terminus 2 agent
76.4%
xhigh
-
Toolathlon-Verified
72.8
xhigh
#10 of 22 models with a reported result
-
VideoMMEv2
61.5
xhigh
-
WorldVQA (w/ tools)
49.4
xhigh; w/ tools
-
ZeroBench (w/ tools)
50
xhigh; w/ tools
#1 of 2 models with a reported result
Independent results so far
46 scores from 46 boards that test models themselves. Ranks compare Muse Spark Open with every model on the same board.
Coding (13)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| SWE-Bench Pro V2 | 95.0% | #12 of 17 | mini-swe-agent, xhigh | Scale SEAL leaderboards |
| SciCode | 58.4% | #17 of 109 | xHigh | Artificial Analysis |
| ALE-Bench | 1590.1 | #25 of 112 | xHigh | ALE-Bench (Sakana AI) |
| LiveBench coding | 55.1% | #30 of 34 | — | LiveBench |
| FrontierCode 2.0 | 7.6% | #44 of 59 | xHigh | Artificial Analysis |
| Vibe Code Bench (Vals AI) | 76.2% | #44 of 109 | OpenHands | Vals AI boards |
| GameBench 2 | 59.1 | #47 of 95 | xHigh | DuelLab GameBench 2 |
| Code migration (Vals AI) | 21.2% | #50 of 75 | — | Vals AI boards |
Show all 13
| LMArena WebDev | 1527 | #51 of 106 | Max | LMArena WebDev |
|---|---|---|---|---|
| LMArena WebDev: React | 1525 | #51 of 98 | Max | LMArena (Hugging Face dataset) |
| Terminal-Bench 4 (Vals AI) | 6.6% | #52 of 61 | Mini-SWE-agent | Vals AI boards |
| Terminal-Bench 4.0 (Artificial Analysis run) | 8.1% | #52 of 115 | xHigh | Artificial Analysis |
| LMArena WebDev: HTML | 1480 | #62 of 105 | Max | LMArena (Hugging Face dataset) |
Agentic work (9)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Harvey's Legal Agent Benchmark (Vals AI) | 25.0% | #2 of 38 | — | Vals AI boards |
| Finance Agent (Vals AI) | 61.3% | #3 of 74 | — | Vals AI boards |
| LiveBench research | 82.7% | #17 of 34 | — | LiveBench |
| LiveBench browser | 26.2% | #28 of 34 | — | LiveBench |
| LiveBench long agentic | 12.1% | #31 of 34 | — | LiveBench |
| Legal Research Bench (Vals AI) | 38.5% | #39 of 87 | — | Vals AI boards |
| Tax Agent Bench (Vals AI) | 26.3% | #41 of 81 | — | Vals AI boards |
| Excel Modeling Benchmark (Vals AI) | 53.7% | #49 of 72 | — | Vals AI boards |
Show all 9
| APEX-Agents | 33.4% | #53 of 60 | — | Epoch AI benchmarking hub |
|---|
Reasoning (8)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Humanity's Last Exam (Scale run) | 28.0% | #12 of 19 | — | Scale SEAL leaderboards |
| HLE-Diamond (Artificial Analysis run) | 26.3% | #27 of 78 | xHigh | Artificial Analysis |
| Humanity's Last Exam (Artificial Analysis run) | 43.9% | #30 of 186 | xHigh | Artificial Analysis |
| LiveBench reasoning | 73.7% | #33 of 34 | — | LiveBench |
| ARC-AGI | 89.0% | #35 of 75 | xHigh | Epoch AI benchmarking hub |
| CritPt v2, hard subset (Artificial Analysis run) | 29.5% | #36 of 60 | xHigh | Artificial Analysis |
| ARC-AGI-2 | 57.5% | #40 of 76 | xHigh | Epoch AI benchmarking hub |
| GPQA Diamond (Artificial Analysis run) | 89.0% | #49 of 166 | xHigh | Artificial Analysis |
General (6)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| MultiChallenge | 74.2% | #5 of 25 | — | Scale SEAL leaderboards |
| SAGE (Vals AI) | 85.8% | #14 of 28 | — | Vals AI boards |
| LiveBench data analysis | 60.2% | #30 of 34 | — | LiveBench |
| Vals Index | 48.6% | #35 of 38 | — | Vals Index |
| GDPval-AA | 1401 | #46 of 118 | xHigh | Artificial Analysis |
| Long context reasoning (AA-LCR) | 78.3% | #61 of 175 | xHigh | Artificial Analysis |
Web design (6)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| Design Arena: websites | 1302 | #16 of 135 | — | Design Arena |
| Design Arena: UI components | 1300 | #20 of 125 | — | Design Arena |
| Design Arena: all code categories | 1294 | #21 of 130 | — | Design Arena |
| LiveBench design | 58.1% | #32 of 34 | — | LiveBench |
| Design Arena: game dev | 1255 | #36 of 129 | — | Design Arena |
| Design Arena: 3D | 1229 | #44 of 122 | — | Design Arena |
Intelligence (2)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| AA Intelligence Index | 24.9 | #39 of 63 | — | Artificial Analysis (via OpenRouter) |
| Epoch Capabilities Index | 155.4 | #41 of 126 | — | Epoch AI benchmarking hub |
Knowledge (1)
| Benchmark | Score | Rank | Setting | Source |
|---|---|---|---|---|
| AA-Omniscience | 25.8 | #17 of 62 | xHigh | Artificial Analysis |