Meta Launches Muse Spark Open for Agentic Coding on NVIDIA GB300 NVL72

Meta Launches Muse Spark Open for Agentic Coding on NVIDIA GB300 NVL72 Meta has released Muse Spark Open, an 816-billion-parameter mixture-of-experts model tailored for agentic coding and multimodal reasoning, optimized for NVIDIA GB300 NVL72 hardware. Muse Spark Open features an 816B total parameter MoE architecture with 42B active parameters per token and a 1M token context window. A unified 1.8B perception encoder natively processes interleaved text, images, video frames, and audio in a single pass. Using NVFP4 checkpoints on NVIDIA GB300 NVL72, the model reaches up to 50K tokens per second per GPU and 600 tokens per second per user. Deployment and training are supported via SGLang, vLLM, TensorRT-LLM, NVIDIA Dynamo, NeMo AutoModel, and NVIDIA NIM. Architecture and Multimodal Agentic Workflows Muse Spark Open is designed for sustained, multistep agentic software engineering rather than single-shot completions. The model activates 42B parameters across eight of its 256 experts per token and maintains direction using goal conditioning and context compaction across a 1M token context window. In addition to planning, tool calling, and self-correction, it can browse the web, coordinate subagents, and generate explanatory commentary prior to actions. Its integrated 1.8B perception encoder allows the model to analyze text, images, error logs, audio narration, and video walkthroughs concurrently without separate modality pipelines. Performance on NVIDIA GB300 NVL72 The model utilizes NVFP4 checkpoints to lower memory footprint and leverage native FP4 Tensor Core acceleration on Blackwell Ultra GPUs. The GB300 NVL72 platform links 72 GPUs across an NVLink domain delivering 130 TB/s of bandwidth, mitigating interconnect bottlenecks for MoE expert routing. When running agentic workloads with speculative decoding, the setup achieves over 50K tokens per second per GPU and over 600 tokens per user, providing lower latency and expanded KV cache capacity for concurrent long-context sessions. Deployment, Routing, and Fine-Tuning Tooling For serving, NVIDIA supports Muse Spark Open through SGLang, vLLM, and TensorRT-LLM. NVIDIA Dynamo offers deployment recipes that separate compute-bound prefill from decode across distinct GPU pools and applies KV cache-aware routing to eliminate redundant precomputation across turns. Post-training is supported via NVIDIA NeMo AutoModel for direct PyTorch SFT and LoRA fine-tuning using Hugging Face weights, with NVIDIA Megatron Bridge available for large-scale Megatron conversion. The model's NVFP4 weights are available on Hugging Face, alongside deployment containers via NVIDIA NIM.

Meta发布面向智能体编程的Muse Spark Open模型，适配NVIDIA GB300 NVL72平台 Meta推出专为智能体编程优化的816B混合专家（MoE）模型Muse Spark Open。该模型具备42B激活参数和1M上下文窗口，基于NVIDIA GB300 NVL72优化，提供多模态感知与长程代码理解能力。 Muse Spark Open拥有816B总参数与42B激活参数，配备1M token上下文窗口及原生NVFP4检查点支持。 集成了1.8B原生多模态感知编码器，可统一处理文本、图像、视频帧和音频，并具备计算机使用与子代理协调功能。 在具备130 TB/s NVLink带宽的NVIDIA GB300 NVL72上，实现单GPU每秒达50K token、单用户每秒达600 token的高吞吐推理。 支持SGLang、vLLM、TensorRT-LLM及NIM容器化部署，后训练可通过NeMo AutoModel与Megatron Bridge执行。 专为智能体编程构建的多模态架构 Muse Spark Open采用816B MoE架构，通过从256个专家中选取8个的Top-8路由机制，每个token仅激活42B参数。其1M token上下文窗口能够容纳包含复杂依赖与数千个文件的完整项目。模型内置1.8B感知编码器，无需独立流水线即可统一处理文本、图像、视频帧及音频，适用于结合屏幕录制、架构图及错误日志的排障场景。此外，它具备跨桌面、移动端及浏览器的计算机使用能力，可在执行动作前输出说明，并支持多步骤自我纠错与子智能体拆分协同。 NVIDIA GB300 NVL72平台上的推理性能 该模型在NVIDIA GB300 NVL72 Blackwell Ultra平台上进行训练与优化，其72块GPU构成的NVLink域提供高达130 TB/s的双向通信带宽，消除了跨网络传输专家流量的瓶颈。借助原生FP4 Tensor Core加速与Meta生成的NVFP4检查点，内存占用显著降低，释放出更多容量用于长上下文会话的KV缓存。在智能体工作负载下，配合推测解码技术，单GPU吞吐量可达每秒50K token以上，单用户吞吐量超过每秒600 token。 灵活的部署选项与后训练生态 在部署方面，开发者可选用SGLang、vLLM与TensorRT-LLM，配合NVIDIA Dynamo实现Prefill与Decode计算池的分离，并利用感知KV缓存的路由避免重复计算；企业还可通过NVIDIA NIM进行一键式容器化部署。针对后训练，开发者可以使用NVIDIA NeMo AutoModel直接对Hugging Face格式权重进行全量SFT或LoRA微调，亦可通过NVIDIA Megatron Bridge转换为Megatron格式以支持大规模生产级训练。

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. The 1M token context window holds an entire project, including complex dependencies, legacy code, and thousands of files without breaking a task into pieces. As a reasoning model, Muse Spark Open adds audio and video perception alongside image and document understanding, working natively and reliably across what it sees, hears, and reads. NVFP4 checkpoints generated by Meta and optimized for NVIDIA Blackwell Ultra shrink the memory footprint of all 816B parameters with native FP4 Tensor Core acceleration. Fewer bytes moved per token to increase tokens per second per user, and the freed-up memory supports larger KV cache for concurrent long context sessions, delivering higher throughput, a lower cost per token. Custom-built for agentic coding multimodal understanding across text, image, video, and audio General coding models are optimized for single-shot completions such as generating a function, writing a unit test, or suggesting an inline fix. They excel at discrete, well-scoped tasks where the context is small, and the answer is immediate. Agentic coding demands something fundamentally different. Planning a multifile refactor, navigating thousands of files, recovering from failing tests, and coordinating subagents over long horizons requires a model built for sustained, multistep work. Muse Spark Open is built for exactly these workloads, delivering higher first-attempt accuracy, reliable tool calling, and deep codebase understanding across repositories. It plans, acts, observes, and self corrects. It supports multistep task execution with tool use, output evaluation, and iterative refinement. The model uses goal conditioning to hold direction and context compaction to retain the knowledge needed to sustain progress across long-running sessions. With a 1M token context window, it holds an entire project, including complex dependencies, legacy code, and thousands of files without breaking tasks into pieces. Beyond coding, it searches the web (including multimodal browsing), spawns and coordinates subagents for decomposable tasks, and leverages computer use capabilities unified across desktop, mobile, and browser. It generates separate commentary before each action, so developers know what the agent is doing and why. The model’s 1.8B perception encoder natively processes interleaved text, images, video frames, and audio without separate pipelines per modality. This means that the model reasons over what it sees, hears, and reads in a single unified pass. This suits developer workflows spanning screenshots, diagrams, error logs, and video walkthroughs within the same agent session. Sparse MoE routing activates only 42B of 816B total parameters per token, delivering frontier agentic capability at a fraction of the serving cost of dense models at equivalent scale. Figure 1 . Muse Spark Open sparse MoE architecture activates 42B of 816B parameters per token through top 8 routing across 256 experts Video 1. Muse Spark Open diagnoses a bug from a screen recording with narrated audio, reading the codebase to identify the root cause Muse Spark Open optimized performance on GB300 NVL72 The NVIDIA GB300 NVL72 features a rack scale architecture that integrates 72 NVIDIA Blackwell Ultra GPUs into a single platform. Its large, 72-GPU NVIDIA NVLink domain enables efficient communication at 130 TB/s, eliminating bottlenecks that appear when expert traffic must cross traditional off-the-shelf networks. For a sparse MoE model like Muse Spark Open, which routes each token to eight of 256 experts, that bandwidth directly determines how much of the theoretical throughput a deployment realizes. Agentic workloads turn static single-turn interactions into multistep workflows that reason, invoke tools, coordinate subagents, and carry growing context from one turn to the next. Out of the box, Muse Spark Open running agentic workloads on NVIDIA GB300 NVL72 delivers up to 50K per second per GPU and up to 600 tokens per second per user, enabling AI factories to run large parameter models in production at high throughput and low latency. F igure 2. Running agentic workloads, Muse Spark Open on four NVIDIA Blackwell GPUs with speculative decode achieved over 50K tokens per second per GPU and over 600 tokens per user Flexible deployment options for Muse Spark Open NVIDIA supports multiple inference stacks to meet a variety of developer needs. SGLang, vLLM , and TensorRT LLM give developers direct control over serving configuration, parallelism, and batching on the NVIDIA accelerated computing platform. TensorRT LLM adds compiled engines tuned for NVIDIA GPUs. NVIDIA Dynamo publishes deployment recipes for all three, and separates prefill from decode across dedicated GPU pools, so compute bound prompt processing doesn’t slow token generation. Its KV cache-aware routing sends each request to the worker already holding the relevant prefix, avoiding precomputation as context accumulates across long agentic sessions. Post-train Muse Spark Open with NeMo AutoModel and Megatron Bridge Developers can post-train Muse Spark Open on domain specific datasets using NVIDIA NeMo AutoModel , a PyTorch native training library with native Hugging Face checkpoint support and no model conversion required. It enables full SFT and LoRA fine tuning out of the box, optimized for rapid experimentation over a wide variety of industry specific cases. Developers can also leverage NVIDIA Megatron Bridge to convert Hugging Face checkpoints to Megatron format when they are ready for large production-scale training to take advantage of the optimizations. Get started with GPU-accelerated Muse Spark Open To get started, download Muse Spark Open NVFP4 weights from Hugging Face. For production deployments, try the downloadable NVIDIA NIM . The NIM reduces setup to a single command, replacing benchmarking and tuning cycles with a container that automatically selects the best serving backend and precision for the detected hardware.