# Marin 535B-A23B reaches training halfway mark as kernel updates lift MFU to 27%

Percy Liang and the Marin Community team have reached the halfway mark in training the open Marin 535B-A23B model on CoreWeave. Better kernels have lifted MFU from 21% to 27%, with heldout loss tracking forecasts built from four preliminary runs.

Language: en
Time zone: America/Los_Angeles

HTML: https://didcodexreset.com/news/e70937b1ee5242fc8d32d143.html

Content language: en
Localization state: sameLanguage

Source: Weights & Biases · Published 10/7/2026, 11:38:52

Percy Liang announced that training for the open Marin 535B-A23B model has crossed the halfway mark of its hero run. The model is being developed by Liang and the Marin Community team on CoreWeave infrastructure, with all run changes and metrics tracked publicly in Weights & Biases.

According to Weights & Biases, the run's heldout loss aligns with projections derived from four smaller runs conducted prior to the main training run. Kernel optimizations have improved MFU from 21% to 27%. Run metrics and change logs are documented in the [Weights & Biases training report](https://forge.coreweave.com/wandb/marin-community/marin_moe/reports/535B-A23B-18T-Token-Hero-Run-Scaling-Ladder--VmlldzoxNzc2MDM5Ng).

Tags: Marin, Open-Source AI, Weights & Biases, CoreWeave

[View original post](https://x.com/wandb/status/2107901173869252665)