Marin Team Training Largest Fully Open Model Ever: 535B Params, 18T Tokens, 15% Done
Thom_Wolf · x · 2026-09-04
Andy Konwinski and Percy Liang announced that the Marin team is training the largest fully open model ever: Marin 535B-A23B (MoE, 535B total / 23B active) on 18T tokens, now 15% complete.
Highlights:
- Unprecedented transparency: data, logs, and decisions are all open, with a live public tracker
- Compute funded by the Jen-Hsun and Lori Huang Foundation via CoreWeave; Liang publicly thanked Jensen Huang
- Training detail: stable for 20 hours since reverting to the pooled-wave expert-routing backend after a ragged all-to-all swap hung twice; a gate/router weight-decay PR remains unmerged pending the step-60,000 pinned checkpoint
- Thom Wolf (Hugging Face co-founder) amplified the news
More from Infra
- 24/7 AI livestream costs: ~$3,450/day on fal vs $40,000+/day on Seedance — FinanceYF5 · 2026-09-04
- Open-source voice pipeline adds Smart Turn end-of-turn gate before LLM calls — ivan_digital · 2026-09-04
- vLLM team launches Inferact, powers new HUMAIN-M3 Arabic frontier model's inference — woosuk_k · 2026-09-04
- Pedro Domingos: Double LLM efficiency and you should be worth $100B, given Nvidia's math — pmddomingos · 2026-09-04
- AMD MI355x beats Nvidia B300 on tokens-per-dollar TCO in AgentX, SemiAnalysis says — AnushElangovan · 2026-09-04
- Liquid AI launches Nanos: task-specific 350M-2.6B models that run on-device — JosephJacks_ · 2026-09-04