Back to list
Meta Releases Muse Spark Open: An 816B Open Weights Multimodal Powerhouse Built for Local Agentic Work
Open SourceMeta AIOpen WeightsAgentic AI

Meta Releases Muse Spark Open: An 816B Open Weights Multimodal Powerhouse Built for Local Agentic Work

Meta has officially released Muse Spark Open, heralded as the most capable US open-weight artificial intelligence model to date. Hunted on Product Hunt by KP, the launch introduces an 816-billion-parameter Mixture-of-Experts architecture with 42 billion active parameters per token, coupled with a 1-million-token context window and native multimodal capabilities spanning text, image, audio, and video inputs. Designed explicitly for autonomous agentic workflows and long-horizon software engineering, Muse Spark Open demonstrates breakthrough benchmark results, posting an 89.2 score on MCP Atlas and 76.4 on Terminal Bench 2.1. In a historic technical collaboration with Apple, the model is engineered to run locally across two M5 Ultra Mac Studios using the MLX framework. Available immediately on Hugging Face under the Muse Model License and supported across leading cloud platforms, the release marks a pivotal moment for desktop-grade frontier intelligence.

Product Hunt

Key Takeaways

  • Massive Sparse Architecture: Muse Spark Open features an 816-billion total parameter Mixture-of-Experts (MoE) architecture with 42 billion active parameters per token and an expansive 1-million-token context window.
  • Native Multimodal Processing: Unlike models relying on disparate adapters, Muse Spark Open natively processes text, image, video, and audio inputs to generate structured text responses.
  • Local Desktop Frontier Execution: Through direct engineering collaboration between Meta and Apple, the full model can run entirely on-device across two M5 Ultra Mac Studios via Apple's MLX framework.
  • Engineered for Coding Agents: Specifically fine-tuned for multi-step reasoning, tool integration, and autonomous error recovery, scoring 89.2 on MCP Atlas and 76.4 on Terminal Bench 2.1.
  • Hybrid Open Distribution: Released under the Muse Model License with open weights on Hugging Face alongside hosted inference on platforms including Together AI, Fireworks, Groq, and OpenRouter.

In-Depth Analysis

Sparse Mixture-of-Experts Architecture and Multimodal Ingestion

Muse Spark Open represents a major architectural milestone in Meta's model family. Operating as a 62-layer transformer utilizing 256 routed experts with top-8 expert activation per token, the model delivers the operational compute speed of a 42-billion-parameter network while preserving the reasoning capacity and contextual depth of an 816-billion-parameter system. A dedicated 1.8-billion-parameter perception encoder handles image and video tokenization, while a parallel audio pipeline feeds directly into the unified attention backbone. This native multimodal ingestion enables developers to feed extensive codebases, visual UI layouts, system terminal outputs, and live audiovisual sessions across a 1-million-token context window without losing state or coherence.

To accommodate diverse deployment requirements, Meta released checkpoints in both BF16 precision (occupying roughly 1.6 TB) and an NVFP4 quantized checkpoint (approximately 530 GB). The NVFP4 variant is explicitly optimized for next-generation hardware environments, drastically shrinking the memory footprint and enabling native 4-bit tensor core acceleration while maintaining mathematical fidelity during complex code refactoring.

Desktop Frontier AI: The Meta and Apple Engineering Alliance

One of the most consequential aspects of the launch hunted by KP on Product Hunt is the unprecedented hardware collaboration between Meta and Apple. Traditionally, models exceeding several hundred billion parameters have been confined strictly to hyperscale data center clusters running high-bandwidth GPU fabrics. Meta engineers worked directly with Apple's machine learning teams to configure Muse Spark Open across Apple's unified memory architecture.

Using Apple's MLX open-source framework, two M5 Ultra Mac Studios—each providing 512 GB of unified memory—link together to act as a unified desktop compute node with over 1 terabyte of memory bandwidth. In practical tests, the combined setup successfully executed end-to-end autonomous agent tasks on-device. Real-world demonstrations highlighted by Meta leadership included autonomous agents parsing open-source repositories to generate and compile pull requests for the Mastodon client Ice Cubes, as well as orchestrating multi-step automated timeline editing workflows inside DaVinci Resolve without sending a single packet over an external network.

Benchmark Performance in Agentic and Coding Workflows

Rather than targeting conversational evaluation or conventional trivia benchmarks, Meta optimized Muse Spark Open specifically for long-horizon agentic workflows, repository-scale software development, and iterative debugging. On standard agent evaluation suites, the model posted notable results: an 89.2 score on MCP Atlas and a 76.4 on Terminal Bench 2.1.

These metrics underscore the model's reliability in handling continuous tool interaction, maintaining state across bash terminals, diagnosing compile errors, reading API documentation, and autonomously revising codebases when tests fail. The model card explicitly positions Muse Spark Open as a robust foundational engine for multi-agent frameworks, code assistants, and self-directing software engineering systems that require multi-turn autonomy.

Ecosystem Availability and the Muse Model License

Meta has coordinated day-one availability across the AI ecosystem. Weights and model files were uploaded directly to Hugging Face, while inference runtime support was confirmed for SGLang, vLLM, llama.cpp, and Unsloth. Developers seeking serverless API access can immediately run inference across multiple provider networks, including Together AI, Fireworks AI, Groq, Baseten, Modal, Nebius, and OpenRouter, as well as cloud model catalogs like AWS SageMaker, Google Cloud Model Garden, and Microsoft Azure Foundry.

From a governance perspective, the model is governed by the Muse Model License. While it grants free research and commercial implementation for smaller organizations, it establishes specific parameters: commercial products utilizing the model must clearly display "Built with Muse Spark Open" within their interfaces. Furthermore, enterprises generating over $20 million in gross annual revenue offering model-as-a-service commercial deployments must negotiate a separate commercial license agreement with Meta.

Industry Impact

The arrival of Muse Spark Open introduces significant implications across the global AI landscape:

  1. Redefining the Open-Weight Ceiling: By releasing an 816B MoE model that matches or surpasses closed and international open-weight alternatives in agentic coding, Meta reaffirms the competitive viability of open models against proprietary cloud APIs.
  2. Democratization of On-Premise Frontier Workflows: The technical breakthrough enabling 800B+ parameter inference across linked desktop workstations eliminates the cloud dependency for enterprise code security, enabling regulated industries and defense contractors to deploy top-tier coding intelligence entirely within private boundaries.
  3. Acceleration of Agentic Engineering: The transition from conversational AI to autonomous agentic systems requires reliable tool-calling and resilience against cascading errors. Muse Spark Open establishes an accessible baseline for developers building autonomous programming agents, CI/CD triage tools, and complex automation pipelines.

Frequently Asked Questions

What are the core hardware requirements to run Muse Spark Open locally?

To run the unquantized or lightly quantized model locally, users require approximately 1TB of combined unified memory or VRAM. Meta demonstrated local execution using two Apple Mac Studios powered by M5 Ultra processors configured with 512GB of unified memory each, running Apple's MLX framework. On standard server hardware, high-density GPU clusters running NVFP4 or BF16 across vLLM or SGLang are recommended.

What licensing terms govern the commercial use of Muse Spark Open?

Muse Spark Open is released under the Muse Model License rather than permissive licenses like Apache 2.0 or MIT. It allows free adoption for most developers and businesses, provided that user interfaces prominently feature "Built with Muse Spark Open". However, commercial entities exceeding $20 million in annual gross revenue that wish to offer the software as a model-as-a-service must obtain a separate written commercial agreement from Meta.

How does Muse Spark Open handle agentic tasks compared to traditional LLMs?

Unlike traditional large language models optimized for single-turn dialogue, Muse Spark Open was specifically post-trained on long-horizon software engineering tasks. It incorporates a 1-million-token context window and specialized tool-calling capabilities that allow it to read file hierarchies, invoke shell commands, interpret compiler errors, and iterate toward functional solutions across extended workflows.

Related News

Univer Office SDK Brings Editing, Calculation, Collaboration, and File Processing to AI Office Products
Open Source

Univer Office SDK Brings Editing, Calculation, Collaboration, and File Processing to AI Office Products

A GitHub Trending entry from dream-num presents Univer Office SDK as a way to build an AI office inside a product. The description names four capabilities: editing, calculation, collaboration, and file processing. It positions those capabilities for both people and AI agents and says the SDK is built on an extensible open-source core. The central distinction is that Univer is presented as a toolkit for adding office functionality to a product, rather than as a finished office application. The brief entry does not describe how the capabilities are implemented, which file formats are supported, or how AI agents interact with the SDK.

Claude-BugHunter: A Comprehensive Red-Teaming Skill Bundle for Claude Code
Open Source

Claude-BugHunter: A Comprehensive Red-Teaming Skill Bundle for Claude Code

Claude-BugHunter, a trending GitHub project by elementalsouls, introduces a robust skill bundle tailored for Claude Code, specifically designed for bug hunting and external red-team operations. This toolset streamlines security assessments by integrating 82 distinct skills, 15 slash commands, and an extensive library of 681 disclosed-report patterns spanning 24 core vulnerability classes. Furthermore, the bundle enhances security testing by incorporating enterprise identity and infrastructure attack matrices, providing users with a structured framework for complex security analysis within the Claude Code ecosystem.

Enhancing Financial Analysis: Introducing Claude Code Skills for Equity Investors and Traders
Open Source

Enhancing Financial Analysis: Introducing Claude Code Skills for Equity Investors and Traders

Discover the newly trending 'claude-trading-skills' repository on GitHub, designed to enhance financial workflows for equity investors and traders. This toolset integrates Claude Code capabilities to streamline market analysis, technical charting, economic calendar monitoring, stock screening, and the development of trading strategies.