Marin 535B Training Starts with Full Transparency on FLOPs and Configs
_ScottCondron · x · 2026-08-23
The Marin initiative has launched its largest open model training run to date: 535B parameters (Active 23B). The project is notable for extreme transparency:
- Compute Stack: 11 x GB200 NVL72 systems.
- Voyage Plan: 80% pretraining + 20% midtraining on 18.75T tokens, expected to last 3 months, totaling 2.7e24 FLOPs.
- Transparency: Visualizations of domain mixture ratios, sampled documents, live training loss, configs, and scaling laws are publicly accessible.
Prior to the hero run, the team trained a 4-rung scaling ladder (1.6B to 27.7B) to debug issues and forecast performance.
More from Infra
- Windows installer guide for ComfyUI Trellis 2 on RDNA3/3.5/4 GPUs — Wake_Up_Morty · 2026-08-23
- Open Source Project pgrok: Self-Hosted ngrok Alternative via VPS — tom_doerr · 2026-08-23
- Tension between data center opposition and AI industry expansion — NathanpmYoung · 2026-08-23
- Upgrading RTX A6000 thermal paste and fan makes it usable for workloads — cephaloform · 2026-08-23
- Optimized llama.cpp fork for AMD GFX906 (Mi50, Mi60, Radeon VII) — milpster · 2026-08-23
- Nvidia AI Server Prices to Rise 15%+, GB300s Around $600k — zephyr_z9 · 2026-08-23