Mimir v1: A 1B-Parameter HRM Model Built Entirely on Permissible Data

A new open language model called Mimir v1 has appeared on arXiv, and at 1 billion parameters on the Hierarchical Reasoning Model (HRM) backbone, it's the kind of release where we want to actually read the paper — not just the abstract.

According to the authors, Mimir v1 was trained from scratch using only permissible post-training data. For practitioners, that's the real headline, not the parameter count. The training-data constraint determines what this model is good for, what we can legally deploy it on, and what we need to verify before trusting the weights in production.

What the paper actually delivers

Mimir v1 is a 1B-parameter model built on the HRM architecture, trained from scratch rather than fine-tuned from an existing checkpoint. The defining choice is the post-training pipeline: only data the authors classify as permissible was used. That's a governance decision as much as a technical one, and it sets this release apart from typical open-weights drops that rely on whatever corpora happen to be available.

We want to flag three things to watch for when the official code lands:

• The exact data filtering pipeline. The word "permissible" is doing heavy lifting in the abstract, and the implementation matters for anyone attempting to reproduce the results.
• Training configuration and tokenizer details — the boilerplate items that always surface as gotchas during downstream fine-tuning.
• Independent benchmark runs on domains outside the paper's own evaluation suite.

Where this fits in your stack

Open 1B models sit in a useful zone — small enough to fine-tune on commodity hardware, large enough to hold structured reasoning behavior. If the performance claims hold up under independent evaluation, Mimir v1 becomes a legitimate candidate for replacing closed-weight baselines in latency-sensitive or privacy-sensitive pipelines where the training-data trail needs to be auditable end-to-end.

Before you adopt, run the usual sanity checks: evaluate on your own domain data rather than relying solely on the published leaderboard, confirm the license supports your use case, and verify the context window matches your existing serving wrapper. The paper is the starting point for our work — not the finish line.