Inference
EN
Meta

Muse Spark Open

Muse Spark Open is a reasoning model from Meta. 31 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Reasoning
  • Input: text · image · video · file · audio
  • Released 2026-10
RankedEnough results for an index.
IndexAll categories, out of 100.65.1±3.9Median 51.0
CoverageIndex weight that has results.90%
SpeedOutput tokens per second.86/sMedian 67/s
First tokenSeconds to the first output token.0.5 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$1.25Median $0.60
Output / 1MUS dollars per 1M output tokens.$4.25Median $2.55
ContextMaximum tokens in one request.1.05MMedian 262K
EloLMArena rating and rank.—Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
meta/muse-spark-open
Max outputMaximum output tokens per request.
262K
Knowledge cutoffLast date of training data.
—
ToolsTool calls on OpenRouter.
Yes
JSON outputOutput follows a given JSON schema.
Yes
Cache read / 1MUS dollars per 1M cached tokens.
$0.15
WeightsLink to the published weights.
Hugging Face (opens in a new tab) ↗
RetiresDate OpenRouter removes the model.
—

Where to run

APIs that serve this model.
5 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
Fireworks AIHost$1.25$4.251.05M
ModalHost$1.25$4.251.05M
NebiusHost$1.25$4.251.05M
OpenRouterRouter$1.25$4.251.05M
Together AIHost$1.25$4.251.05M

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Run locally

822.04B parameters
16-bit
1727.3 GB
8-bit
918.1 GB
4-bit
524.3 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.9 counted64.846 / 218
CodingCode writing and repair.5 counted63.352 / 209
ReasoningLogic problems and puzzles.6 counted62.356 / 227
MultimodalTasks with images and text.4 counted69.813 / 84
KnowledgeFacts and expert knowledge.4 counted65.443 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.None counted——
MathMath problems.3 counted64.930 / 168

Results

31 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
MCP AtlasAgentic89.2%1 / 52
ZeroBenchMultimodalAbove fitFar above this model's index.50.0%1 / 10
Artificial Analysis Humanity's Last ExamKnowledge43.9%30 / 222
Artificial Analysis MMMU-ProMultimodal80.9%16 / 113
Artificial Analysis SciCodeCoding58.4%17 / 114
OTIS Mock AIME 2024-2025Math98.1% ±1.331 / 200
BabyVisionMultimodal80.2%2 / 11
SimpleVQAMultimodal73.6%4 / 18
JobBenchAgentic61.8%8 / 36
Humanity's Last Exam with toolsAgentic61.0%8 / 36
Critical Physics TasksReasoning14.9%51 / 214
GPQA diamond · Epoch AIKnowledge88.6% ±1.952 / 217
Artificial Analysis GPQA DiamondKnowledge89.0%48 / 199
SimpleQA VerifiedKnowledge56.6% ±1.622 / 91
Artificial Analysis Omniscience AccuracyKnowledge42.0%44 / 164
Artificial Analysis Long Context ReasoningReasoning78.3%63 / 217
GDPval-AA normalizedAgentic45.1%39 / 133
ARC-AGI-1 (semi-private)Reasoning89.0%39 / 126
LiveBench ResearchAgentic80.3%17 / 52
Vibe Code Bench v1.1Coding76.2% ±3.341 / 123
FrontierMath-Tiers-1-3-v2-PrivateMath69.8% ±2.730 / 90
ARC-AGI-2 (semi-private)Reasoning57.5%46 / 131
FrontierMath-Tier-4-v2-PrivateMath31.7% ±7.432 / 72
MazeBench: visual spatial reasoning in a 3D open worldAgentic0.0%23 / 48
Toolathlon-VerifiedAgentic72.8%15 / 23
Code MigrationCoding21.2% ±3.551 / 77
DeepSWEAgentic60.0%41 / 59
LiveBench Coding 2.0Coding73.6%37 / 52
LiveBench Browser UseAgentic36.0%35 / 49
Terminal-Bench 4.0Agentic6.6% ±0.549 / 61
LiveBench Long AgenticAgentic18.9%42 / 51
LiveBench ReasoningReasoning73.7%71 / 80
LiveBench Data AnalysisReasoning60.2%72 / 80
LiveBench DesignAgentic60.3%47 / 52
ARC-AGI-3 (semi-private)Reasoning0.1%24 / 26
FrontierCode 2.0Coding6.2%27 / 27
Finance Agent (v2)Finance61.3% ±2.13 / 79
HLAB-Vals-1.1Legal25.0% ±3.62 / 35
ARC-AGI-1 (public eval)Reasoning94.0%38 / 116
ARC-AGI-2 (public eval)Reasoning57.2%44 / 122
Legal Research BenchLegal38.5% ±3.439 / 90
SAGEEducation85.8% ±1.714 / 29
Tax Agent BenchFinance26.3% ±2.742 / 82
EMBFinance53.7% ±3.251 / 75
SWE-Atlas Codebase QnACoding58.0%—

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseEpoch AICC BYARC Prize FoundationTerms restrict reuseLiveBenchNo licence statedVals AINo licence statedmodels.devMIT

History

5 releases
Index65.1 → 65.1
23 Oct 202612 Nov 2026
Output / 1M$4.25 → $4.25
23 Oct 202612 Nov 2026

Questions

How much does Muse Spark Open cost?
$1.25 per 1M input tokens, $4.25 per 1M output tokens.
How good is Muse Spark Open?
Its index is 65.1 out of 100. That is place 44 of 283 ranked models.
What is Muse Spark Open best at?
Multimodal: 69.8 out of 100, place 13 of 84.
How fast is Muse Spark Open?
86 output tokens per second, 0.5 s to the first token.
What is the context window of Muse Spark Open?
1.05M tokens.
Is Muse Spark Open open source?
Yes. Its weights are published.
When was Muse Spark Open released?
October 2026.

Badge

Muse Spark Open on Inference360: place 44 of 283, index 65.1
HTML
<a href="https://inference360.ai/models/muse-spark-open"><img src="https://inference360.ai/badges/muse-spark-open.svg" alt="Muse Spark Open on Inference360: place 44 of 283, index 65.1" height="20"></a>
Markdown
[![Muse Spark Open on Inference360: place 44 of 283, index 65.1](https://inference360.ai/badges/muse-spark-open.svg)](https://inference360.ai/models/muse-spark-open)

Compare with

Muse Spark 1.264.9Muse Ember64.8GPT-5.6 Luna65.5DeepSeek V4 Pro 081365.5dots3-note Preview63.4

Lower price

dots3-note Preview63.4 · FreeGLM-5.3-Flash63.1 · $0.078/1MLing 3.1 Flash67.2 · $0.135/1M

More from Meta

All models
Muse Wonder Full77.3Muse Wonder73.1Muse Ember64.8Muse Spark 1.3 ContributorUnrankedMuse Spark 1.370.1