# DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data

Source: [arXiv](https://arxiv.org/abs/2608.13517v1)  
Feed7 permalink: https://feed7.dev/p/2608-13517v1-10qer54  
Published: 2026-08-13T17:37:53.000Z  
Trust: Needs Review (needs_review)

## Why Included

Mimir v1 is an open 1B-parameter reasoning model trained with permissible post-training data. It is relevant for compact or Danish deployments, though the abstract supplies no benchmark scores.

## Source Summary

Mimir v1 is a **1B-parameter** language model based on the Hierarchical Reasoning Model architecture and released on Hugging Face. Its training mixture uses **161 datasets**, with permissible post-training data.

## Practical Implication

The authors evaluate it across **20 benchmarks** covering English, math and code, and Danish. Builders considering compact or Danish-language models now have an openly available candidate whose data sourcing is a stated design constraint.

## Agent-Ready Context

Mimir v1 is a **1B-parameter** language model based on the Hierarchical Reasoning Model architecture and released on Hugging Face. Its training mixture uses **161 datasets**, with permissible post-training data.

The authors evaluate it across **20 benchmarks** covering English, math and code, and Danish. Builders considering compact or Danish-language models now have an openly available candidate whose data sourcing is a stated design constraint.

The paper claims gains over HRM-Text 1B and competition with larger Qwen and Gemma models, but the supplied material includes no task-level scores, licenses, latency, or hardware results. Those details need checking before model selection.

## Connected Context

Feed7 judgment across 462 accumulated Signals:

This adds an open, compact model candidate whose permissible-data constraint and Danish evaluation broaden selection criteria beyond English benchmark performance and parameter scale. It supports testing whether a 1B hierarchical architecture can cover reasoning, code, and regional-language needs, but the absence of task scores, licensing details, latency, and hardware results prevents comparison with larger open or hosted alternatives.

- [Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAI](https://feed7.dev/p/data-quality-is-the-compute-multiplier-ari-morcos-datologyai-0x7k2ve) — The data-quality candidate argues that curation and task matching can substitute for some scale; Mimir makes data permissibility and a 161-dataset mixture part of that compact-model proposition, though it supplies insufficient scores to quantify the effect.
- [The Base Model Is Dead — Varun Singh, Arcee AI](https://feed7.dev/p/the-base-model-is-dead-varun-singh-arcee-ai-02hts76) — The base-model discussion says code and reasoning priors increasingly depend on data composition; Mimir provides a concrete small-model case spanning code, math, English, and Danish, but does not reveal which mixture choices produced its claimed gains.
- [Local Models: Trust, Control, Optimization — Carter Abdallah, NVIDIA](https://feed7.dev/p/local-models-trust-control-optimization-carter-abdallah-nvidia-17u7gz9) — Mimir supplies an openly available model for the ownership and customization strategy described by the local-model candidate, while its missing license and serving measurements leave key governance and operational checks unresolved.
- [Inkling Small from Thinking Machines is now available on AI Gateway](https://feed7.dev/p/inkling-small-now-available-on-ai-gateway-1a9781l) — Both invite evaluation of smaller models for capable workloads, but Mimir emphasizes open release, permissible training data, and Danish coverage while Inkling emphasizes routed multimodal and tool-use features; neither supplied record establishes workload-level efficiency parity.

## Context Map

- Layer: model
- Domains: coding
- Topics: open-models, reasoning, model-selection

## Uncertainty

- The paper claims gains over HRM-Text 1B and competition with larger Qwen and Gemma models, but the supplied material includes no task-level scores, licenses, latency, or hardware results. Those details need checking before model selection.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.