Editorial illustration for Meta Muse Spark Open on NVIDIA GB300 NVL72 Scales Agentic Coding Workflows
Meta's 816B Muse Spark Open Powers Agentic Coding
Meta Muse Spark Open on NVIDIA GB300 NVL72 Scales Agentic Coding Workflows
Meta put a number on the table Wednesday: 816 billion parameters, 42 billion of them active at any time, routed through a mixture-of-experts design called Muse Spark Open. The model was trained on NVIDIA's GB300 NVL72 rack-scale system, built around Blackwell Ultra GPUs, and it's aimed squarely at a problem general coding models tend to fumble, which is holding an entire codebase in mind while working through it.
A 1 million token context window is the headline spec. That's enough room to load a full project, dependencies and legacy code and all, without chopping the job into fragments the model has to reassemble later. Meta also shipped NVFP4 checkpoints tuned for Blackwell Ultra, cutting the memory load of those 816 billion parameters and freeing up space for a bigger KV cache, which matters when multiple long-context sessions are running at once.
Most coding models are built for single-shot work, a function here, a unit test there. Agentic coding is a different kind of task entirely, one that spans thousands of files and multiple steps over time.
Meta launches Muse Spark Open, a coding-optimized 816B mixture-of-experts (MoE) model with 42B active parameters and a 1M token context window. Trained on NVIDIA GB300 NVL72 Blackwell Ultra specifically for agentic coding workflows, it delivers higher first attempt accuracy, reliable tool calling, and deep codebase understanding across repositories.
Why this matters
A 1M token context window plus an 816B MoE architecture with only 42B active parameters tells us Meta is chasing a specific problem: agents that lose the thread halfway through a real codebase. That's the actual bottleneck in agentic coding today, not raw parameter count. If Muse Spark Open genuinely holds entire repositories, legacy dependencies included, without chunking tasks apart, that's worth testing against Claude and GPT-based coding agents on messy, multi-file projects rather than toy benchmarks.
The 1.8B perception encoder handling text, images, audio, and video in one pass also matters for teams whose workflows involve screenshots, architecture diagrams, and recorded walkthroughs alongside code, since stitching separate vision and language pipelines together has been a real source of friction. Running this on GB300 NVL72 Blackwell Ultra hardware is a reminder that these capability gains are tied to infrastructure most teams don't own, which raises the real question: will this ship as something developers can run themselves, or stay locked to NVIDIA's top-tier clusters. We'd want independent benchmarks on tool-calling reliability before trusting any first-attempt accuracy claims.
Related Reading
Browse all Meta news →
AI Tools & Apps Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps OpenAI Developer: AI Agent "Swarms" Waste Tokens, Create "Coordination Tax
AI Tools & Apps LG's recent webOS update adds Microsoft Copilot app, now removable
Related Topic Meta Faces Lawsuit Over AI Training Data, Facial Recognition
Policy & Regulation
Related Topic OpenAI hires 630 ex-Meta staff as ChatGPT memory may turn data into ads
Business & Startups
From Archives CopilotKit Releases CLiF to Convert Agent Traces Into Fine-Tuning Data
From Archives K2K explores NVIDIA AI factories for Liverpool City Region public services
From Archives Meta Muse Wonder vs Claude: Rival Runs Best-Effort, Not Common Harness
Business & Startups
From Archives Meta Unveils Muse Wonder AI Models for Code Subscribers
Business & Startups
Common Questions Answered
What is the mixture-of-experts design in Meta's Muse Spark Open and how does it differ from standard large language models?
Muse Spark Open uses a mixture-of-experts (MoE) architecture with 816 billion total parameters, but only 42 billion of them are active at any given time. This design allows the model to be more efficient than traditional models with the same parameter count, as it routes computations through specialized expert networks rather than activating all parameters simultaneously.
How does the 1 million token context window of Muse Spark Open address limitations in agentic coding workflows?
The 1 million token context window enables Muse Spark Open to load and maintain entire codebases in memory while working through them, solving a key bottleneck where coding agents typically lose context halfway through real projects. This allows the model to understand legacy dependencies and multi-file projects without needing to chunk tasks apart or lose track of the overall codebase structure.
What hardware infrastructure was used to train Meta's Muse Spark Open model?
Muse Spark Open was trained on NVIDIA's GB300 NVL72 rack-scale system, which is built around Blackwell Ultra GPUs. This specialized hardware infrastructure was specifically chosen to support the training of this large mixture-of-experts coding model.
What specific coding problems is Muse Spark Open designed to solve compared to general coding models?
Muse Spark Open is specifically optimized for agentic coding workflows and addresses the problem of holding an entire codebase in mind while working through it, which general coding models tend to struggle with. The model delivers higher first attempt accuracy, reliable tool calling, and deep codebase understanding across repositories, making it better suited for complex, multi-file development tasks than traditional coding models.
Further Reading
- Scale Agentic Coding Workflows with Meta's Muse Spark Open on NVIDIA GB300 NVL72 - NVIDIA Developer Blog
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta - Hugging Face
- Meta ships Muse Spark Open, its answer to China's best open-weight models - The New Stack
- Meta and Apple team up to make Muse AI models run better on Macs - The Verge
- Meta launches new Muse Spark Open model, optimized to run on Mac Studio - 9to5Mac
