Intel / article

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

Source: developer.nvidia.com
Author: Tanya Lenz
Date: Published Oct 15, 2026

Context

Meta announced Muse Spark Open as an open-weight model for agentic coding. This NVIDIA post covers the serving side: what it takes to run a model of 816B parameters quickly enough for long agent sessions. Muse Spark Open is a model, meaning each token is routed to 8 of its 256 expert sub-networks, so only about 42B parameters do work on any given token. That keeps the computation per token well below a dense model of the same size, but every one of the 816B parameters still has to be held in memory, and moving each token's work to experts spread across many GPUs can become the bottleneck.

NVIDIA pairs two answers to that. Its GB300 NVL72 rack joins 72 Blackwell Ultra GPUs into a single NVLink domain running at 130 TB/s, so expert traffic does not have to cross a slower general-purpose network; NVIDIA says that bandwidth decides how much of the model's theoretical throughput a deployment actually gets. Meta also produced checkpoints in NVFP4, a 4-bit number format that Blackwell GPUs compute in natively. Smaller weights mean fewer bytes moved per token, which raises per-user speed, and the memory they free can hold a larger , the stored context of each session, so more long sessions against the 1M-token window can run at once.

The headline figures are up to 50K tokens per second per GPU and up to 600 tokens per second per user. The chart behind them says it was measured on four Blackwell GPUs with , a technique in which drafted tokens are checked in batches, so these are best-case numbers for that setup rather than a general guarantee. For teams that want to run it themselves, NVIDIA points to SGLang, vLLM and TensorRT LLM with Dynamo deployment recipes, a downloadable NIM container that picks a serving backend for the detected hardware, and NeMo tools for fine-tuning on domain data.

Key takeaways · AI-distilled

Muse Spark Open has a 1M token context window, which the post says can hold an entire project with dependencies and thousands of files without splitting the task.
The model uses goal conditioning and to stay on track in long sessions, and it writes separate commentary before each action so developers can see what the agent is doing.
NVIDIA says the GB300 NVL72 links 72 Blackwell Ultra GPUs in one NVLink domain at 130 TB/s, which matters because routing each token to 8 of 256 experts generates heavy expert traffic.
Dynamo recipes cover SGLang, vLLM, and TensorRT LLM, splitting prefill and decode onto separate GPU pools and routing requests to workers that already hold the relevant KV cache prefix.
Fine-tuning is supported through NeMo AutoModel for full SFT and LoRA directly on Hugging Face checkpoints, with Megatron Bridge converting to Megatron format for larger production training.

Terms in this piece · Glossary

mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
KV cache — The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.

Read the source developer.nvidia.com