Marin

Stanford CRFM's "open lab": fully-open models (weights, data, code, tokenizer; Apache-2.0) pretrained from scratch in JAX on TPUs via the Levanter trainer, with every experiment conducted as a public GitHub PR. Marin 8B (May 2025) was the first fully-open-process 8B of its kind; Marin 32B Base (Dec 2025) beats OLMo 2 32B Base on 14 of 19 standard benchmarks — the largest fully-open-process university-trained model to date.