---
title: "Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72"
description: "Meta launches Muse Spark Open, an 816B-parameter mixture-of-experts model with 42B active parameters and a 1M token context window, trained on NVIDIA GB300 NVL72 for agentic coding workflows. It plans, acts, observes, and self-corrects across multistep tasks like multifile refactoring, using tool calls and subagent coordination while holding an entire project in context without breaking it into pieces. Its 1.8B perception encoder natively processes interleaved text, images, video frames, and audio in a single pass. On GB300 NVL72 with speculative decoding, it achieves over 50K tokens per second per GPU and over 600 tokens per second per user. Sparse MoE routing activates only 42B parameters per token, and NVFP4 quantization reduces memory footprint for larger KV cache. The model is available via Hugging Face and NVIDIA NIM, with support from SGLang, vLLM, and TensorRT LLM for serving, and fine-tuning through NeMo AutoModel and Megatron Bridge."
---

# Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72

[NVIDIA](https://yomu.fyi/company/nvidia-developer-blog) · Tanya Lenz · Oct 15, 2026

**Type:** Announcement

## Summary

Meta launches Muse Spark Open, an 816B-parameter mixture-of-experts model with 42B active parameters and a 1M token context window, trained on NVIDIA GB300 NVL72 for agentic coding workflows. It plans, acts, observes, and self-corrects across multistep tasks like multifile refactoring, using tool calls and subagent coordination while holding an entire project in context without breaking it into pieces. Its 1.8B perception encoder natively processes interleaved text, images, video frames, and audio in a single pass. On GB300 NVL72 with speculative decoding, it achieves over 50K tokens per second per GPU and over 600 tokens per second per user. Sparse MoE routing activates only 42B parameters per token, and NVFP4 quantization reduces memory footprint for larger KV cache. The model is available via Hugging Face and NVIDIA NIM, with support from SGLang, vLLM, and TensorRT LLM for serving, and fine-tuning through NeMo AutoModel and Megatron Bridge.

## Context

Traditional coding models optimized for single-shot completions struggle with sustained multistep agentic workflows that require planning, tool use, and long context across thousands of files.

## Approach / What changed

Muse Spark Open uses a sparse MoE architecture with 256 experts, activating the top 8 per token for 42B active parameters out of 816B total. It has a 1M token context window and a 1.8B multimodal perception encoder for text, images, video, and audio. Training and optimization on NVIDIA GB300 NVL72 Blackwell Ultra enable NVFP4 quantization with native FP4 Tensor Core acceleration. Deployment uses SGLang, vLLM, or TensorRT LLM, with NVIDIA Dynamo for separation of prefill and decode and KV cache-aware routing. Post-training fine-tuning is supported via NeMo AutoModel (SFT and LoRA) and Megatron Bridge for multi-node scaling.

## Takeaways

- Muse Spark Open activates only 42B of 816B parameters per token through top-8 routing across 256 experts, reducing serving cost while maintaining frontier agentic capability.
- On NVIDIA GB300 NVL72 with speculative decoding, Muse Spark Open achieves over 50K tokens per second per GPU and over 600 tokens per second per user, enabling high-throughput production deployment for agentic workflows.
- Its 1.8B perception encoder processes text, images, video frames, and audio natively in a single pass, allowing unified reasoning over screenshots, diagrams, error logs, and video walkthroughs within the same agent session.

## FAQ

### How does Muse Spark Open reduce memory footprint without losing quality?

Meta provides NVFP4-weighted checkpoints that leverage native FP4 Tensor Core acceleration on Blackwell GPUs, shrinking the memory footprint of all 816B parameters. Fewer bytes per token increase throughput, and freed memory supports a larger KV cache for longer context sessions.

### Which inference frameworks are recommended for deploying Muse Spark Open?

NVIDIA supports SGLang, vLLM, and TensorRT LLM. TensorRT LLM provides compiled engines tuned for NVIDIA GPUs. NVIDIA Dynamo separates prefill and decode across dedicated GPU pools and uses KV cache-aware routing to avoid precomputation during long agentic sessions.

### How can I fine-tune Muse Spark Open on my own data?

Use NeMo AutoModel for SFT and LoRA fine-tuning with native Hugging Face checkpoint support, no model conversion required. For multi-node scaling, Megatron Bridge adds tensor, pipeline, and expert parallelism without reconverting the checkpoint.

**Tags:** [Agentic Development](https://yomu.fyi/topic/agentic-development), [Coding Agents](https://yomu.fyi/topic/coding-agents), [LLMs](https://yomu.fyi/topic/llm), [Mixture of Experts](https://yomu.fyi/topic/mixture-of-experts), [NVIDIA](https://yomu.fyi/topic/nvidia)

- Source: [NVIDIA](https://developer.nvidia.com/blog/scale-agentic-coding-workflows-with-metas-muse-spark-open-on-nvidia-gb300-nvl72)
- Source URL: https://developer.nvidia.com/blog/scale-agentic-coding-workflows-with-metas-muse-spark-open-on-nvidia-gb300-nvl72
- Ingested by Yomu: 2026-10-15T18:00:21.185Z

[Read original post](https://developer.nvidia.com/blog/scale-agentic-coding-workflows-with-metas-muse-spark-open-on-nvidia-gb300-nvl72)