Passer au contenu principal

Muse Spark Open (Xhigh)

MetaOpen WeightMuse Model License · Usage Commercial

Description

Muse Spark Open is Meta Superintelligence Labs' open-weight natively multimodal Mixture-of-Experts model (816B total / 42B active parameters) with a 1,048,576-token context window. The HF model card specifies text, image, video, and audio inputs and text output. It supports configurable reasoning_effort (low/medium/high/xhigh; default xhigh) and is positioned for coding and agentic workflows. Weights are released under the Muse Model License with limited commercial terms and are available on Hugging Face as meta-models/Muse-Spark-Open-816B-A42B (BF16) and an NVFP4 sibling. Hosted inference is offered by third-party providers including Together AI, Fireworks AI, Groq, Modal, Nebius, Baseten, and OpenRouter (slug meta/muse-spark-open).

Date de sortie
2026-10-15
Paramètres
816.0B
Longueur du contexte
1.0M
Modalités
audio, image, pdf, text, video

Radar de capacités

40
general
58
coding
89
reasoning
69
science
80
agents
70
multimodal

Classements

Domaine#RangScoreSource
Capacité agentique81
44.0
LS
Classement codage45
88.0
AA
Classement général78
66.0
AA
Classement multimodal23
62.0
LS
Raisonnement45
74.0
LB
Science26
83.0
AA

Scores de benchmarks (LLM Stats)

(LLM Stats (zeroeval))

Agents

Toolathlon-Verified72.8%Aut.
Job Bench61.8%Aut.

Code

DeepSWE 1.160.0%Aut.
SWE Atlas - Codebase QnA58.0%Aut.

General

GDPval-AA1401.00 / 3000Aut.
Artificial Analysis39.0%Aut.

Knowledge

AA-Omniscience Index25.00 / 200Aut.

Multimodal

OSWorld-Verified80.4%Aut.

Reasoning

MCP Atlas89.2%Aut.
GPQANYU + Cohere + Anthropic (2023)89.0%Aut.
Terminal-Bench 2.176.4%Aut.
Humanity's Last Exam (with tools)61.0%Aut.
SciCode58.5%Aut.

Vision

BabyVision80.2%Aut.
SimpleVQA0.74 / 100Aut.
ZEROBench0.50 / 100Aut.
WorldVQA49.4%Aut.

Indices d'évaluation AA

(Artificial Analysis)
Gpqa(NYU + Cohere + Anthropic (2023))
89.0
Lcr(Artificial Analysis)
78.3
Terminalbench V2 1
76.4
Scicode(UIUC + Argonne National Lab (2024))
58.4
Hle(Center for AI Safety + Scale AI (2025))
43.9
Intelligence Index(Artificial Analysis)
38.9
Terminalbench V4 0
8.1

Scores par catégorie LLM Stats

(LLM Stats (zeroeval))
Legal
100
Finance
100
Agents
100
Reasoning
100
General
85
Knowledge
25
Science
25
Tool Calling
80
Image To Text
70
Multimodal
70
Physics
70
Biology
70
Chemistry
70
Code
70
Vision
70
Math
60
Productivity
60

Tarification

Prix d'entrée$1.25 / 1M tokens
Prix de sortie$4.25 / 1M tokens
Prix mixte (3:1)$2 / 1M tokens
Prix de lecture cache$0.15 / 1M tokens

Vitesse

Tokens/sec221.9
Délai du premier token0.52s
Temps de réponse9.53s

Classement des Prix par Fournisseur

Classement des Prix par Fournisseur

12 fournisseurs

Moins cher: TogetherPlus cher: Venice AI
FournisseurEntréeSortie
1TogetherMoins cher
$0
$0
2MetaPRINCIPAL
$1.25
$4.25
3NanoGPT
$1.25
$4.25
4OpenRouter
$1.25
$4.25
5Kilo Gateway
$1.25
$4.25
6Cloudflare AI Gateway
$1.25
$4.25
7Ollama Cloud
$1.25
$4.25
8Requesty
$1.25
$4.25
9Vercel AI Gateway
$1.25
$4.25
10Merge Gateway
$1.25
$4.25
11Prime Inference
$1.25
$4.25
12Venice AI
$1.5625
$5.3125

Comparer les prix entre différents fournisseurs API pour ce modèle.

Sources externes