Meta Unveils Muse Spark Open, an 816B Parameter Mixture-of-Experts Agent Model
Meta has released open weights for Muse Spark Open, an 816B-parameter multimodal model with a 1M token context window built for coding and complex agent workflows.
Meta Platforms Inc. has released the open weights for Muse Spark Open, an artificial intelligence system engineered specifically for software development and long-horizon agent tasks. Previously kept proprietary following its initial appearance in April, the model represents what Meta calls the most capable open-weight architecture produced by an American laboratory to date. The system employs a sparse mixture-of-experts structure, activating just 42 billion of its total 816 billion parameters for any individual token.
Muse Spark Open accepts multimodal inputs, allowing users to combine text with video, audio, or images inside a context window that extends up to 1 million tokens. The model's foundation rests on Muse Spark 1.2, which launched in August. To allow the model to operate securely outside of Meta's proprietary hosted environment, the company applied specialized safety post-training. Meta noted that the model satisfied the criteria outlined in its internal Meta Superintelligence Scaling Framework, with an in-depth evaluation paper slated for later publication.
Architecture and Real-World Testing
According to technical documentation published by Hugging Face Inc., the training data for Muse Spark Open emphasized practical software engineering workflows, including extensive debugging tasks and refactoring efforts across multiple files. Beyond standard data center evaluations, Meta also demonstrated the system on local hardware.
In one testing scenario, researchers ran the model across two Apple Inc. Mac Studio computers powered by M5 Ultra processors, connected into a single unified setup via Apple's MLX framework. Equipped with roughly one terabyte of unified memory across both machines, the configuration ran Muse Code to implement new features and resolve issues within Ice Cubes, an open-source Mastodon client. In another evaluation, the system operated within DaVinci Resolve, analyzing raw video on a frame-by-frame basis to identify and crop footage suitable for vertical video formats.
Benchmark Results and Model Context Protocol
Evaluation benchmarks published by Meta show varied strengths depending on the task. On the JobBench agent evaluation, Muse Spark Open attained a score of 61.8, significantly outpacing rival American open models such as Thinking Machines Lab Inc.'s Inkling, which scored 31.1, and Nvidia Corp.'s Nemotron 3 Ultra at 27.2.
In raw coding benchmarks, open models developed in China retained an advantage. On the Terminal-Bench 2.1 coding benchmark, Muse Spark Open posted a score of 76.4, finishing behind Moonshot AI's Kimi K3 (which led with 85), as well as GLM 5.3 and Qwen3.8.
However, Muse Spark Open outperformed all tested competitors on tool integration. On the MCP-Atlas benchmark, the model scored 89.2, surpassing both domestic and Chinese alternatives. In that evaluation, the system received instructions in natural language without explicit tool assignments, requiring the model to autonomously determine the sequence and selection of tools across 36 active Model Context Protocol servers covering 220 total tools.
Licensing Terms and Distribution
Muse Spark Open is issued under the Muse Model License, which includes specific commercial stipulations. Developers can modify the model and embed it within commercial software products without incurring licensing fees. However, organizations providing the model as a hosted service that generate more than $20 million in gross revenue over a 12-month timeframe must negotiate a separate commercial agreement with Meta. Products utilizing the model must display a "Built with Muse Spark Open" notice, and deployments within military operations or critical infrastructure are strictly prohibited.
Meta previously published Muse Glimmer, a 30 billion-parameter model, in August. The company's highest-tier technology, including the Muse Wonder line unveiled on Monday, remains closed-source and accessible exclusively via Muse Code and the Meta Model API. For Muse Spark Open, weight downloads are accessible via Hugging Face and cloud catalogs from Google LLC, Microsoft Corp., and Amazon Web Services Inc., with hosted inference available through seven providers at launch, including Fireworks AI Inc. and Groq Inc.
What it means for developers
For engineering teams building autonomous agents, Muse Spark Open provides an open alternative designed around complex tool execution. Its top performance on the MCP-Atlas benchmark demonstrates strong aptitude for navigating complex Model Context Protocol ecosystems, making it a viable orchestrator for external APIs and specialized developer utilities.
At the same time, deploying an 816-billion-parameter architecture presents obvious operational challenges. Even with an active parameter footprint of 42 billion, the total model footprint requires significant hardware resources—such as multi-machine unified memory clusters or high-end cloud instances—to host locally. For teams evaluating agentic pipelines without maintaining large on-premise hardware, developers can try top AI models cheaply through one API at https://apixoai.online to benchmark latency, cost, and execution quality across different platforms.
Teams must also weigh benchmark nuances: while Muse Spark Open leads in agentic orchestration and multi-tool planning, developers prioritizing pure terminal code generation may find higher accuracy in models like Kimi K3 or closed alternatives. Additionally, commercial distributors approaching Meta's $20 million revenue threshold must factor licensing terms into their long-term infrastructure roadmap.
Source: Meta releases 816B-parameter Muse Spark Open for coding and agent work — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
