Meta has released Muse Spark Open, an 816-billion-parameter mixture-of-experts (MoE) model optimized for coding workflows. The model uses sparse routing to activate 42 billion parameters per token and is distributed with NVFP4 checkpoints generated by Meta and tuned for NVIDIA Blackwell Ultra hardware.
Purpose and significance
Muse Spark Open is designed for sustained, agentic coding workloads rather than one-shot code completions. It targets tasks such as planning multifile refactors, navigating thousands of files, recovering from failing tests, and coordinating subagents across long horizons. The model plans, acts, observes and self-corrects, supporting tool use, output evaluation and iterative refinement.
A key capability is the 1 million–token context window, which can hold an entire project — including complex dependencies, legacy code and thousands of files — without the need to split the task into smaller pieces.
Multimodal perception and unified processing
The model includes a 1.8 billion–parameter perception encoder that natively processes interleaved text, images, video frames and audio in a single pass. This allows the model to reason over what it sees, hears and reads together — useful when developer workflows include screenshots, diagrams, error logs and video walkthroughs within the same agent session.
Muse Spark Open’s sparse MoE architecture uses top-8 routing across 256 experts, so only 42 billion of the total 816 billion parameters are active per token. This approach reduces serving cost compared to dense models of equivalent scale.
Performance on NVIDIA GB300 NVL72
Muse Spark Open has been optimized for the NVIDIA GB300 NVL72. The GB300 NVL72 is a rack-scale system that integrates 72 NVIDIA Blackwell Ultra GPUs into a single platform and provides a 130 TB/s NVLink domain for efficient intra-system communication. For sparse MoE models that route each token to eight of 256 experts, this bandwidth is a determining factor in realized throughput.
According to Meta and NVIDIA, agentic workloads running Muse Spark Open on GB300 NVL72 can reach up to 50,000 tokens per second per GPU and up to 600 tokens per second per user. Those figures aim to enable production deployments of large-parameter models with high throughput and low latency.
Memory and efficiency: NVFP4
NVFP4 checkpoints, produced by Meta and optimized for NVIDIA Blackwell Ultra, reduce the memory footprint of the model’s 816 billion parameters and leverage native FP4 Tensor Core acceleration. Moving fewer bytes per token increases tokens per second per user and frees memory for a larger KV cache in concurrent long-context sessions, yielding higher throughput and a lower cost per token.
Tooling, web browsing and subagents
Muse Spark Open can search the web with multimodal browsing, spawn and coordinate subagents for decomposable tasks, and operate unified computer-use capabilities across desktop, mobile and browser environments. It generates separate commentary before each action so developers can see what the agent intends to do and why.
Deployment and fine-tuning options in the NVIDIA ecosystem
NVIDIA supports multiple inference stacks for Muse Spark Open: SGLang, vLLM and TensorRT LLM, allowing developers direct control over serving configuration, parallelism and batching on NVIDIA accelerated hardware. TensorRT LLM provides compiled engines tuned for NVIDIA GPUs.
NVIDIA Dynamo publishes deployment recipes for all three stacks and separates prefill from decode across dedicated GPU pools. KV cache–aware routing sends each request to the worker already holding the relevant prefix, avoiding redundant precomputation as context accumulates across long agentic sessions.
Developers can post-train Muse Spark Open on domain-specific datasets using NVIDIA NeMo AutoModel, a PyTorch-native training library with Hugging Face checkpoint support and built-in SFT and LoRA fine-tuning capabilities. When workloads exceed a single node, NVIDIA Megatron Bridge takes the same Hugging Face checkpoints and adds tensor, pipeline and expert parallelism so teams can scale from experimentation to production training without reconverting the model.
Getting started with GPU-accelerated Muse Spark Open
Muse Spark Open NVFP4 weights are available for download on Hugging Face. For production deployments, NVIDIA provides a downloadable NIM (NVIDIA Inference Module) that reduces setup to a single command: the container automatically selects the best serving backend and precision for the detected hardware.
Summary
Muse Spark Open is a large, sparsely routed MoE model built for multimodal, agentic coding and long-context tasks. With NVFP4 checkpoints and optimizations for NVIDIA GB300 NVL72, the release targets high throughput and lower token costs in production environments.
