Meta Unveils Gigawatt-Scale Prometheus Supercluster as Muse Wonder Secures #6 on WebDev Code Arena

Background and the Problem

The progression of frontier foundation models and agentic workflows has firmly entered the era of gigawatt-scale infrastructure and massive reinforcement learning. For clusters deploying hundreds of thousands of GPU accelerators, the preeminent operational challenge is "goodput degradation" caused by hardware failures, silent data corruption, and network desynchronization. Frequent node failures trigger expensive checkpoint restores, wasting millions of dollars in compute capital. Meta officially pulled back the curtain on its flagship Prometheus Supercluster and shared milestone results on Muse Wonder across global developer arenas.

Architecture and How It Works

1. Gigawatt-Scale Prometheus Infrastructure (New Albany, Ohio):

• Houses hundreds of thousands of top-tier GPUs interconnected via an ultra-high-bandwidth fabric;
• Delivers sustained 85%+ goodput on pretraining and 90%+ goodput on reinforcement learning pipelines;
• Achieved approximately 47,000 GPU-hours mean time between failures (MTBF) during the training run of Muse Wonder, setting a benchmark for ultra-large distributed system stability;

1. Next-Generation Agentic Foundation: Muse Wonder:

• Architected specifically for long-horizon multi-step planning, tool orchestration, and full-stack web software synthesis;
• Complemented by Muse Glimmer 30B, an open-weight 30B parameter model engineered to run in 4-bit precision on a single consumer GPU or desktop workstation.

Benchmarks and Measured Results

According to LMSYS official announcements and Meta technical disclosures (Figure 1, Arena WebDev, Meta Telemetry):

• LMSYS Code Arena: WebDev Blind Pairwise Elo:
• Muse Wonder Full debuts with 1726 Elo points, securing #6 globally;
• Decisively surpasses Anthropic's Claude Sonnet 5.5 (+16 pts) and Google's Gemini 4 Argon (+48 pts);
• Catapults Meta to #3 among all AI research organizations worldwide, surpassing Google;
• Full-Stack Web Engineering Capabilities:
• Demonstrates superior compilation integrity on dynamic UI component synthesis, multi-route state machines, and modern enterprise frameworks (React 19, Next.js, and modern styling libraries).

Getting Started for Developers

• Live Testing: Developers can immediately test and benchmark Muse Wonder in Code Arena: WebDev on arena.lmsys.org;
• On-Device Agent Strategy: Meta confirmed upcoming open weights for Muse Glimmer 30B, enabling developers to run self-hosted 30B agent pipelines within 16GB–24GB of VRAM for local, privacy-preserving coding assistants.