Marin

weights

open(Apache-2.0)

data

open(documented mixtures)

code

open(marin + Levanter, JAX)

checkpoints

open

reproducibility

bit-for-bit

license

Apache-2.0(OSI)

Marin publishes Apache-2.0 weights together with its training code, data mixtures, experiments, hyperparameters and training logs, and its training runs are reproducible bit for bit. Nothing in the stack is withheld, so Marin is a fully open model.

• http://marin.community/blog/2025/05/19/announcement/ recorded 2026-09-29

  Announcement post: Marin is an open lab "in which the research and development of models is completely transparent from day 1"; each experiment is tracked by a GitHub issue, and "Everything is reproducible, declared in code, and open to the community to see."

• https://api.github.com/repos/marin-community/marin recorded 2026-08-13

  `private: false`, `archived: false`, license `Apache-2.0`, description "Open-source framework for the research and development of foundation models", last pushed 2026-08-13.

• https://developers.googleblog.com/stanfords-marin-foundation-model-first-fully-open-model-developed-using-jax/ recorded 2026-09-29

  Google Developers post: "the models, data, code, and tokenizer are all released under the permissive Apache 2.0 license", and the stack kept "bit-for-bit reproducibility across preemptions" while migrating TPU slices.

• https://huggingface.co/api/models/marin-community/marin-32b-base recorded 2026-09-29

  Hugging Face API for marin-community/marin-32b-base: gated false, tag "license:apache-2.0", dataset tags including "dataset:common-pile/stackv2_edu_filtered".

• https://huggingface.co/api/models/marin-community/marin-8b-base recorded 2026-09-29

  Hugging Face API for marin-community/marin-8b-base: gated false, private false, tag "license:apache-2.0", and dataset tags including "dataset:mlfoundations/dclm-baseline-1.0".

• https://huggingface.co/marin-community/marin-32b-base/raw/main/README.md recorded 2026-09-29

  Marin 32B Base card: "The code and model are released under Apache 2.0."; repositories "Core repo (data and experiment management): https://github.com/marin-community/marin" and Levanter training code; each training phase links a GitHub issue and experiment script.

• https://huggingface.co/marin-community/marin-8b-base recorded 2026-09-29

  Marin 8B Base model page, tagged "license:apache-2.0" with four safetensors weight shards and not gated; the card links a "Core repo (data and experiment management):" and Levanter training code.

• https://huggingface.co/marin-community/marin-8b-base/raw/main/README.md recorded 2026-09-29

  Raw Marin 8B Base card: front matter "license: apache-2.0" with the dataset list (dolmino-mix-1124, olmo-mix-1124, dclm-baseline-1.0, starcoderdata, proof-pile-2 and Marin markdownified sets); "The code and model are released under Apache 2.0." Checkpoint table lists main at 12.7T tokens plus intermediate branches.

Adoption is measured as Hugging Face downloads summed across Marin's two declared artifacts, the 8B base and instruct checkpoints. It is a research-stage, fully-open project included on the map as an example of full openness rather than for its popularity.

• https://huggingface.co/api/models/marin-community/marin-8b-base recorded 2026-10-07

  898 downloads in the trailing 30 days for marin-community/marin-8b-base

• https://huggingface.co/api/models/marin-community/marin-8b-instruct recorded 2026-10-07

  438 downloads in the trailing 30 days for marin-community/marin-8b-instruct

On its own 19-task evaluation suite, Marin 32B Base scores about level with Gemma 3 27B and ahead of OLMo 2 32B, but trails Qwen 2.5 32B. The project has released no post-trained 32B or reasoning model to set against OLMo 3.1 Think, which keeps it a step below OLMo.

• http://marin.community/blog/2025/05/19/announcement/ recorded 2026-09-29

  Announcement post: "14 out of 19 standard base model evals (MMLU, HellaSwag, GPQA, etc.), Marin 8B Base outperforms Llama 3.1 8B Base"; Marin 8B Instruct "outperforms OLMo 2 on standard instruct model evals, but still fall short of".

• https://developers.googleblog.com/stanfords-marin-foundation-model-first-fully-open-model-developed-using-jax/ recorded 2026-09-29

  Google Developers post: "The first releases from this open lab are the" Marin-8B-Base and Marin-8B-Instruct models; the Tootsie process "spanned over 12 trillion tokens".

• https://huggingface.co/marin-community/marin-32b-base/raw/main/README.md recorded 2026-09-29

  Marin 32B Base card, LM Eval Harness table: "| **Marin 32B (Mantis)** | 65.2 |" average against "| OLMo 2 32B Base | 63.2 |", "| Gemma 3 27B PT | 65.1 |" and "| Qwen 2.5 32B Base | 68.1 |".