Muse Spark 1.3
MultimodalMuse Spark 1.3 is Meta Superintelligence Labs' proprietary multimodal reasoning model for long-running agentic, multi-agent, and coding workflows. It improves on Muse Spark 1.2 in long-horizon collaboration, multitasking, tool use, failure recovery, and concise coding execution, supports a 1M-token context window, and is available through Muse Code and the Meta Model API at $0.10 per million input tokens.
Key Specifications
Parameters
-
Context
1.0M
Release Date
September 2, 2026
Average Score
73.4%
Timeline
Key dates in the model's history
Announcement
September 2, 2026
Last Update
September 3, 2026
Technical Specifications
Parameters
-
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval
Pricing & Availability
Input (per 1M tokens)
$0.10
Output (per 1M tokens)
$0.20
Max Input Tokens
1.0M
Max Output Tokens
943.7K
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning
Benchmark Results
Model performance metrics across various tests and benchmarks
Other Tests
Specialized benchmarks
Agentic IF Index (Internal)
Meta internal composite instruction-following evaluation for agentic workflows. • Self-reported
AutomationBench
AutomationBench public v3 task set; deterministic end-state grading; pass@1. • Self-reported
DeepSearchQA
Official 900-question dataset with a common browser harness; F1 score. • Self-reported
DeepSWE 1.1
DeepSWE v1.1 official 113-task set with mini-swe-agent; max reasoning; task pass rate. • Self-reported
GDPval-AA
GDPval-AA v2 Elo in the Artificial Analysis Stirrup agentic harness; human baseline 1,000 (max_score 3000). • Self-reported
Job Bench
Official 65-task set with a file-aware rubric grader; mean rubric score. • Self-reported
MRCR v2 (8-needle)
OpenAI MRCR v2 8-needle, 256K-512K context range; 100 examples; mean sequence-matcher ratio. • Self-reported
MRCR v2 (8-needle, 512K-1M)
OpenAI MRCR v2 8-needle, 512K-1M context range; 100 examples; mean sequence-matcher ratio. • Self-reported
OSWorld 2.0
OSWorld 2.0 version 08.08 in Meta's common internal computer-use evaluation framework; mean per-task partial score. • Self-reported
SWE Atlas - Codebase QnA
SWE-Atlas public Codebase QnA split with mini-swe-agent and task-specific rubric grading; mean pass@1. • Self-reported
Terminal-Bench 2.1
Terminal-Bench 2.1 official 89-task set in an isolated cloud sandbox; mean pass@1. • Self-reported
License & Metadata
License
proprietary
Announcement Date
September 2, 2026
Last Updated
September 3, 2026
Compare Muse Spark 1.3
All comparisonsArticles about Muse Spark 1.3
Similar Models
All ModelsMuse Spark 1.1
Meta
MM
Released:Jul 2026
Price:$1.25/1M tokens
Muse Spark 1.2
Meta
MM
Released:Aug 2026
Price:$0.10/1M tokens
Muse Spark
Meta
MM
Best score:0.9 (TAU)
Released:Apr 2026
Llama 3.2 90B Instruct
Meta
MM90.0B
Best score:0.9 (MMLU)
Released:Sep 2024
Price:$1.20/1M tokens
Muse Glimmer-30B
Meta
MM29.6B
Best score:0.8 (GPQA)
Released:Aug 2026
Llama 3.2 11B Instruct
Meta
MM10.6B
Best score:0.7 (MMLU)
Released:Sep 2024
Price:$0.18/1M tokens
Llama 4 Scout
Meta
MM109.0B
Best score:0.8 (MMLU)
Released:Apr 2025
Price:$0.18/1M tokens
Llama 4 Maverick
Meta
MM400.0B
Best score:0.9 (MMLU)
Released:Apr 2025
Price:$0.27/1M tokens
Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.