Meta’s Muse Spark Open: A New Open-Weight Agentic Model

At a Glance
| Item | Value |
|---|---|
| Publisher | Together AI |
| Published | 2026-10-15 |
| Source type | Unverified (not confirmed by a primary source) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
It is reported that Meta’s new open-weight model, “Muse Spark Open," has become available on Together AI’s platform. However, this information has not yet been widely confirmed by official Meta channels and is considered unconfirmed information based on an announcement by the cloud platform Together AI.
According to a Together AI blog post, Muse Spark Open was developed for long-running agentic coding workflows that operate while keeping the codebase of an entire repository within context. It is reported to be built for computer operation and web automation across desktops, mobiles, and browsers, acting as an autonomous agent in software engineering, as well as the integrated understanding of documents, images, audio, and video.
Specifications
Specifications confirmed from materials published by Together AI are as follows. Note that detailed breakdown parameters of the architecture are not explicitly stated in the text of the material, but the model identifier on the API is confirmed as meta-models/Muse-Spark-Open-816B-A42B.
- Context length: 1M tokens
- Input modalities: Text, images, audio, audio-video
- Operating characteristics: Autonomous planning, execution, observation, self-correction, and cooperative control of sub-agents
- Audit support function: Intent commentary, which outputs what action will be taken and why before executing it
It is explained that the vast context window of 1 million tokens allows entire projects containing older code with complex dependencies or thousands of files to be fed whole into the model without splitting them up by task.
Performance
At this time, third-party verification and specific benchmark test score lists (comparison tables) have not been published. However, qualitative performance evaluations and feedback from early access customers are introduced in Together AI’s blog post.
According to the publisher Together AI, Muse Spark Open is positioned as “currently the most capable US-made open-weight model." It is claimed that by using actual development work data from the field actually performed by software engineers, such as multi-file refactoring, debugging sessions, and long-running work, it achieves improved accuracy on the first attempt, reliable tool calling, and a deep codebase understanding capability.
Evaluations have also been submitted by companies that participated in the early access program prior to its official release. Zimu Li from Factory, a provider of software development agents, reported that in agent engineering workflows, Muse Spark Open consistently selected appropriate tools and maintained concise, lean responses required for production-grade agents. Furthermore, Dong Chen, a senior AI engineer at Ollama, a provider of local LLM execution environments, evaluated that it showed particularly high performance in tool calling and long-context processing, generating clean structured outputs even in multi-step and parallel calls.
Because these evaluations remain subjective and qualitative reports from the publisher and early partners, objective capabilities and the degree of superiority over existing major models await future publication of benchmark scores and community verification.
Strengths and Use Cases
It is reported that Muse Spark Open is designed with a focus on utilization as an autonomously operating software engineering agent, rather than stopping at mere text generation or code completion. Based on Together AI’s published content, the main use cases and areas of strength assumed for this model are as follows:
Agentic Development Targeting Large-Scale Codebases
It is stated that this model uses data from practical tasks performed in actual development sites, such as multi-file refactoring, long debugging sessions, and complex fix operations, for training. By leveraging the vast context window of up to 1 million tokens, it is possible to load complex dependencies, legacy code, or entire repositories on the scale of thousands of files directly into the context without splitting them by task. This is expected to enable high-accuracy fixes and feature additions from the first attempt without losing the context of the entire project.
Computer Operation and Multimodal Web Automation
It is stated to support computer use across desktop environments, mobile devices, and web browsers, as well as multimodal web browsing. Application is envisioned for workflows that explore the web and autonomously execute various operations while interpreting visual screen information and text information in combination.
Complex Multimodal Understanding
It supports not only text as an input modality, but also images, audio, and audio-video. It is said to be suited for use cases that advance development and investigation while comprehensively interpreting document charts, UI screenshots, and recorded videos or audio logs for operation checks.
Long-Term Autonomous Task Planning, Correction, and Sub-Agent Cooperation
It is equipped with a loop that autonomously plans with a long-term perspective, executes, observes the results, and performs self-correction. Furthermore, for tasks that can be divided, a function to spin up multiple sub-agents and cooperatively control them is also envisioned. Because it features an intent commentary function that verbalizes and outputs “what it will do" and “why it will do it" before execution, a key characteristic is that human developers can easily audit and trace the intermediate decision-making process even during long autonomous execution.
Robust Tool Calling and Structured Output
As emphasized in the evaluations of early access customers, it is said to excel in the ability to accurately execute multi-step tool calls and parallel calls, outputting lean and organized structured data. It is suited for tasks that select appropriate tools and execute them precisely in production-grade workflows, such as Factory’s autonomous development agent “Droids".
How to Get It
At this time, the methods announced by Together AI for using Muse Spark Open are the platform’s cloud API and playground. It is also guided that it will be provided from Ollama, a local LLM execution tool.
Using via Together AI Serverless API
It can be used through Together AI’s Serverless API by specifying the model identifier meta-models/Muse-Spark-Open-816B-A42B. A basic calling example using the Python client library is as follows:
from together import Together
client = Together()
response = client.chat.completions.create(
model="meta-models/Muse-Spark-Open-816B-A42B",
messages=[
{
"role": "user",
"content": "Refactor this module to remove the circular import.",
}
],
)
print(response.choices[0].message.content)
API usage pricing is guided as follows:
- Input tokens: $1.25 per million tokens
- Output tokens: $4.25 per million tokens
- Cached input tokens: $0.15 per million tokens
A playground where you can try out behaviors on a browser without writing code is also provided on Together AI.
Availability on Ollama
In comments from an engineer at early access partner Ollama, it was expressed that the model will be provided to Ollama users from day one through a partnership with Together AI. However, the current announcement materials do not contain specific download commands (such as ollama run) or details of the distribution format. For the local distribution of the open-weight model files (weight data) themselves and details of the license terms, it is necessary to check future official guidance from Meta and each platform.
Related Articles
- Optimizing Open Models for Long-Horizon Agentic Workloads
- Meta Releases Muse Spark Open: Multimodal MoE Model
Sources
This article contains unverified information. We will append an update note once it is confirmed by a primary source.

