Model Releases
Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
Meta’s Muse Spark Open is an 816‑billion‑parameter mixture‑of‑experts coding model that activates only 42 billion parameters per token through top‑8 routing across 256 experts, and has been trained on
Meta’s Muse Spark Open is an 816‑billion‑parameter mixture‑of‑experts coding model that activates only 42 billion parameters per token through top‑8 routing across 256 experts, and has been trained on NVIDIA GB300 NVL72 (Blackwell Ultra). The model offers a one‑million‑token context window, enabling end‑to‑end analysis of an entire codebase—including thousands of files and complex dependencies—along with native audio‑video perception via a 1.8B encoder; it delivers up to 50 k tokens/second per GPU (600 tokens/s/user) through NVFP4 checkpoint optimization that reduces memory usage with FP4 tensor‑core acceleration. Users can download the NVFP4 weights from Hugging Face or deploy a single‑command NVIDIA NIM for production workloads.
Related
- Run Local Agentic AI Workflows with Meta’s Muse Glimmer on NVIDIA
- Experiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Coding
- Muse Spark Open is now on Ollama. @Meta's open-weight model for coding and tool use, with a 1M-token context window. Run it on Ollama’s clou…
- Meta releases Muse Spark Open, saying it is the most capable US open-weight model, with 816B total and 42B active parameters and a 1M-token context window (Meta AI Research)
Source: NVIDIA Developer | 2026-10-15