Marin 535B open-source model training starts with 2.7e24 FLOPs plan
DavidBennett__ · x · 2026-08-24
Percy Liang announced the start of training for the Marin 535B-A23B model. The project is fully open, including code, data, recipes, and experimental results. Training will run on 11 GB200 NVL72 clusters for approximately 3 months, involving a total compute of 2.7e24 FLOPs across 18.75T tokens for pre-training and mid-training. The team previously debugged and forecasted the run using a scaling ladder from 1.6B to 27.7B models.
More from Infra
- Practice: Self-Improving AI Chip Design via RL Hits 87k tps for Kimi K3 — brianryhuang · 2026-08-24
- FLUX.2 Klein image generator ported to pure C/CUDA — camenduru · 2026-08-24
- OpenRouter Reveals Decentralized Compute Providers Including Darkbloom — toptickcrypto · 2026-08-24
- Claude Code Users Report 529 API Overload Errors — backpropagation6 · 2026-08-24
- Aquaduck launches 'Community Cloud' for low-cost inference on home devices — punkyrockypocky · 2026-08-24
- Daily vs Monthly Budgets Shape AI Development Workflows — schwentker · 2026-08-24