Meta's Muse Spark Open served on single TPU v7 host via SGLang and TorchTPU

PyTorch engineers served Meta's Muse Spark Open, an 816B-parameter mixture-of-experts model, on a single TPU v7 host with four chips using SGLang on the TorchTPU backend.

Read the full story at PyTorch Blog ↗

Timeline · 1 report

2026-11-10 16:45 · PyTorch Blog
Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU

More stories

Meta expands AI age estimation globally to apply Instagram Teen Account protections — Meta Newsroom
Meta launches Muse Ember, a low-cost AI model aimed at high-volume agent tasks — TechCrunch AI
Meta's AI agent Muse arrives on Apple Watch and Android Wear OS — TechCrunch AI
AI leaders at Google DeepMind, Meta and OpenAI tease imminent AI-driven scientific breakthroughs — OpenAI YouTube
Meta chief strategy officer Dave Wehner to join AI startup Fireworks as CFO — CNBC Technology
Anthropic has its own AI-powered office suite now — The Verge AI
eBay appears to block third-party AI agents, including Meta's Muse, from browsing its platform — The Verge AI
AI agents see limited traction for ride-hailing and food delivery apps — Business Insider AI

Get the daily brief of stories like this at 6:30 every morning →