One-line: 1B-parameter model trained from scratch on permissible data (HRM-Text architecture, 70.5B tokens/epoch over 161 datasets) — SOTA for Danish, competitive English (~1B class). Apache-2.0. ⚠ small 4K context and Danish/English-only scope limit agent-loop utility.
Coding benchmarks
Benchmark
Score
Source
HumanEval
56.7
lab-reported (card)
HumanEval+
—
not published
MBPP
—
not published
LiveCodeBench
—
not published
Aider polyglot
—
not published
SWE-bench Verified
—
not published
SWE-bench Multilingual
—
not published
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
—
not published
τ-Bench
—
not published
ToolACE
—
not published
GAIA
—
not published
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
57.5 (MMLU)
lab-reported (card)
GPQA Diamond
—
not published
MATH
45.8
lab-reported (card)
AIME 2025
—
not published
GSM8K
89.9
lab-reported (card)
Efficiency
Active params: ~1B dense (no MoE, all active)
Context length: 4,096 tokens
VRAM (fp16 rough): ~2 GB weights
Runtime: standard transformers (model_type hrm_text, safetensors, endpoints_compatible), no custom kernels needed
What makes it notable
DFM-Mimir is a rare open 1B model trained entirely from permissible, audit-scrutinized data (Danish project by Univ. Southern Denmark / Aarhus / Copenhagen + Alexandra Institute, Ministry-funded) with a published memorisation/copyright audit and technical report. It beats prior HRM-Text 1B and sits near larger 2-4B peers on its English benchmark average (69.0 vs Qwen3.5-4B 69.3) and is the clear SOTA for Danish. For the coding/agent tracker specifically it’s a modest candidate: coding ability is a side effect of general training (HumanEval 56.7, MATH 45.8) rather than a focus, and the 4K context plus Danish/English-only scope make it a poor agent-loop workhorse compared to 1B models like MiniCPM5-1B (128K context, native tool calling). Its real-value case is researchers/users who need an open, ethically-sourced, audit-clean small model for Danish-language or low-risk English tasks.