Meta Releases 816-Billion-Parameter Open-Weight AI Model for Coding and Agent Work

Meta Platforms has released Muse Spark Open, an 816-billion-parameter open-weight AI model that the company describes as the most capable open-weight model released by a U.S. laboratory to date. Built for coding and long-running agent workflows, the model features multimodal input capabilities spanning text, image, audio, and video, along with a 1-million-token context window.

The model employs sparse activation technology, utilizing only 42 billion parameters per token despite its massive total parameter count. Built from Meta's Muse Spark 1.2 base version released in August 2026, Muse Spark Open has undergone additional safety post-training to function without the guardrails of Meta's own services. Meta confirmed the model passed release thresholds in its Superintelligence Scaling Framework, with a full evaluation report still to come.

Benchmark Performance Demonstrates Tool-Use Strength

Muse Spark Open demonstrated exceptional performance on tool-use benchmarks, achieving a score of 89.2 on MCP-Atlas—the highest among all tested models, including Chinese competitors. The benchmark presents models with plain-language tasks without naming required tools, leaving the model to determine which of 220 tools across 36 live Model Context Protocol servers to call and in what sequence.

On the JobBench agent test, the model scored 61.8, more than doubling the 31.1 recorded by Inkling from Thinking Machines Lab and significantly outperforming Nvidia Nemotron 3 Ultra at 27.2. However, on Terminal-Bench 2.1 coding assessments, Muse Spark Open achieved 76.4, trailing Chinese models including Kimi K3 (85), GLM 5.3, and Qwen3.8.

According to a Hugging Face technical writeup, the model's training focused on "real developer workflows" including multi-file refactors and extended debugging sessions.

Deployment Options and Desktop Demonstration

Meta demonstrated the model running on two Apple Mac Studio desktops equipped with M5 Ultra chips that Apple's MLX framework linked into one system. The two machines had approximately 1 terabyte of combined memory between them.

During testing, Muse Code ran on this setup while the coding terminal worked through bug fixes and new features for the open-source Mastodon client Ice Cubes. In another demonstration, the model processed footage in DaVinci Resolve frame by frame to identify and crop optimal shots for vertical video.

The model is available through seven inference providers at launch, including Groq Inc. and Fireworks AI Inc., with downloadable weights accessible via Hugging Face, Amazon Web Services, Google LLC, and Microsoft Corp.

Licensing Terms and Restrictions

The model is released under a new Muse Model License carrying what Meta calls "limited commercial terms." No fee applies to developers who adapt the model or embed it within their own software. However, Meta requires a separate agreement from any company that sells the model as a hosted service and has exceeded $20 million in gross revenue over a 12-month period.

All products built on the model must display a "Built with Muse Spark Open" label. Military and critical infrastructure uses are prohibited under the license terms.

Meta's release of Muse Spark Open, substantially larger than its 30-billion-parameter Muse Glimmer model from August 2026, continues the company's strategy of releasing select models as open-weight while maintaining proprietary Muse Wonder models as closed systems accessible only through Meta Model API and Muse Code.

Sources

1. Meta releases 816B-parameter Muse Spark Open for coding and agent work — SiliconANGLE