Tools
marin, an open framework for foundation-model research, is running visible TPU training infrastructure in public at 443 stars a week
marin-community/marin (2,343 stars, Apache-2.0) is unusual in that its commit log is an operating training cluster rather than a library changelog: commits on 2026-08-25 and 08-26 publish levanter checkpoint phase telemetry, add merged cache catalogs, allow the TPU ferry launcher on preemptible capacity, default the TPU ferry to v6e, make GCP IAM role bindings authoritative and repair a vLLM Grafana overview. For builders, the value is less the framework and more the visible reference implementation of what running foundation-model training on preemptible TPU capacity actually requires, including the Grafana dashboards and auto-triage CI.
↳ Follow the thread