Meta Muse Spark 1.3 Release

Meta has released Muse Spark 1.3, a model specifically engineered for long-horizon agentic workflows and high-performance competitive coding. The release focuses on improving first-attempt accuracy, reliable tool calling, and native multimodal perception, positioning it as a competitive alternative to frontier models in coding tasks.

High-Performance Coding and Agentic Workflows

Muse Spark 1.3 is designed to handle complex, multi-step reasoning tasks that require tracking context and prior results over long horizons. Unlike standard LLMs, the model is tuned to work through messy or conflicting inputs and proactively ask for clarification when necessary to reduce unnecessary turns in a workflow.

Key technical capabilities include:

Agentic Optimization: Optimized for long-horizon tasks, enabling the model to act as a a development partner or the core of a coding agent.
Competitive Coding: Tuned for cleaner output and higher first-attempt accuracy, performing competitively with frontier models across several coding evaluations.
Native Multimodal Perception: The model can perceive video, images, and documents. Its visual reasoning is executed through a real execution environment rather than scripted steps, allowing it to build software based on screenshots or video clips.

Benchmarks and Performance

According to community reports and provided data, Muse Spark 1.3 has shown strong results in coding benchmarks. Specifically, it achieved a DeepSWE score of 75.4, which users noted as a top-tier score. Additionally, some users reported that Muse Spark 1.3 Max has surpassed OpenAI's best models on the Artificial Analysis Intelligence Index.

Pricing and Data Privacy Tiers

Meta has introduced a transparent, tiered pricing structure for Muse Spark 1.3 that explicitly links cost to data privacy. The model is available in two primary versions:

Model	Context Window	Input (per Mtok)	Cached Input (per Mtok)	Output (per Mtok)	Data Usage
`muse-spark-1.3-contributor`	1M	$0.10	$0.002	$0.20	Used to improve Meta products
`muse-spark-1.3`	1M	$1.25	$0.15	$4.25	Not used to improve Meta products

The "contributor" tier offers a significant discount (approximately 20x) for users who allow Meta to train on their data, a move that has been praised by some for its transparency and criticized by others as a way to monetize user data.

Developer Ecosystem and Tooling

Meta provides several ways to deploy and integrate Muse Spark 1.3:

Muse Code: A multi-agent coding harness for the terminal, built specifically for Muse Spark.
Meta Model API: A self-serve API that is OpenAI SDK compatible.
OpenRouter: Integration for swapping the model into existing workflows.

Meta has also released several "cookbooks" to demonstrate advanced agentic capabilities, including multi-agent orchestration (creating a SaaS app from a one-line idea), agent fan-out (parallel worktrees), computer use (giving the model "eyes and hands" to build software), and search grounding for real-time web access.

Community Insights and Counterpoints

Developer feedback on Muse Spark 1.3 is mixed but generally positive regarding its speed and cost-efficiency.

"I started using Spark 1.2 for development... it felt like it knew its weaknesses and didn't try to impose its opinions on me... it did what I told it and if there was something unexpected in the code that it put out it was often because I gave it ambiguous or conflicting instructions."

Some users expressed concerns about the model's tendency to overfit on certain paths or that it may not yet be the absolute state-of-the-art (SOTA) for every use case, with some citing Gemini 3.8 Flash or Sol models as superior alternatives in specific scenarios. Others highlighted the ethical concerns regarding Meta's corporate history, which influenced their choice of model provider.