Модель DFM Mimir v1 от датского университета обошла коммерческие аналоги в задачах логического мышления

A 1-billion-parameter language model built at the University of Southern Denmark matched models four times its size on standard reasoning benchmarks — trained without a single token sourced from proprietary or legally contested data, with weights now freely available on Hugging Face. The paper, available on the arXiv preprint server, appeared on August 13, 2026, authored by Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, and Lukas Galke Poech under the Danish Foundation Models (DFM) project. The model, called DFM Mimir v1, is built on the Hierarchical Reasoning Model (HRM) architecture and trained on a mixture of 161 datasets, all drawn from sources the authors classify as permissible for post-training use. That combination — a non-standard architecture and an explicit data-ethics constraint — puts Mimir v1 at the intersection of two of the most contested questions in applied AI: whether parameter-scaling is the only path to frontier-level reasoning, and whether competitive models can be built without the training pipelines that have drawn active copyright litigation against most major AI labs. The answer Mimir v1 offers, provisionally and pending independent replication, is yes on both counts. Read more: Microsoft Build 2026: MAI-Thinking-1 Is First In-House Reasoning Model, Trained Without OpenAI Data Inside the HRM: How Architecture Replaces Scale Standard large language models process tokens through a fixed-depth stack of transformer layers in a single forward pass. The depth is determined at build time and does not change based on how difficult the problem is. Chain-of-thought prompting — the dominant workaround — externalizes reasoning into visible token sequences, which adds latency and fragility while requiring large volumes of reasoning-trace training data. The foundational CoT research by Wei et al. established how much extra scaffolding standard models require to handle multi-step problems at scale. The Hierarchical Reasoning Model, originally proposed by Guan Wang and colleagues at Sapient Intelligence (the original HRM architecture paper) in 2025, takes a structurally different approach. Rather than generating intermediate reasoning tokens, HRM runs two interdependent recurrent modules in a nested loop during a single forward pass: a high-level module responsible for slow, abstract planning, and a low-level module handling rapid, fine-grained computation. The two modules iterate over the same input embeddings across multiple cycles, injecting state additively. A learned halting signal — or a fixed schedule, in HRM-Text's implementation — determines when computation is complete. In HRM-Text 1B, the default configuration is two outer H-module cycles, each with three L-module updates, for eight total recurrent steps per forward pass. This is equivalent, in training-compute terms, to four recursions of a standard model — but those four recursions are spent cycling a fixed set of parameters rather than propagating through additional layers. The architecture provides what the paper calls "effectively unbounded compute depth at bounded parameter count": harder problems can be given more recurrent steps without adding parameters or generating visible reasoning tokens. Training stability in recurrent architectures is notoriously difficult to achieve, because repeated gradient propagation through the same parameters produces multiplicative gradient instability. HRM-Text addressed this with two innovations described in its May 2026 paper: MagicNorm, a normalization scheme that provides PostNorm-like forward stability while behaving like a more gradient-friendly PreNorm scheme during the truncated backward pass; and warmup deep credit assignment, which starts backpropagating through only the last two recurrent steps and gradually expands to five, preventing the gradient spikes that plagued prior recurrent-depth approaches. The second major architectural choice is the training objective. Rather than predicting every next token in raw internet-scale text, HRM-Text trains exclusively on instruction-response pairs, applying loss only over the response tokens. Instruction tokens use full bidirectional attention (as in encoder-decoder architectures); response tokens use standard causal attention. This PrefixLM approach concentrates every gradient update on task completion rather than on predicting prompt-like or task-irrelevant text. The result, according to the HRM-Text paper, is a model that reaches competitive reasoning performance on roughly 40 billion unique tokens — an amount 100 to 900 times smaller than the training datasets used by comparable standard models. One important deployment note: because of the PrefixLM attention pattern, the model performs as benchmarked only when inference matches the training setup — a detail that matters for developers evaluating the model in production inference frameworks. HRM-Text's creators acknowledged that coding performance lags behind reasoning and general-knowledge performance, a limitation that reflects the dataset composition rather than a fundamental architectural barrier. Whether Mimir v1's 161-dataset mixture addresses this gap is a question the new paper does not fully resolve. What Mimir v1 Costs to Train, and Why That Matters HRM-Text 1B, Mimir v1's closest architectural predecessor, was trained in approximately 46 hours on two eight-GPU H100 nodes — 16 GPUs total — at an estimated cost of approximately $1,472. The corresponding training runs for models that compete with it on benchmarks — Qwen 3.5 2B (900 times more tokens), Gemma 3 4B (100 times more estimated compute), Llama 3.2 3B (225 times more tokens) — cost proportionally more. The DFM paper does not separately report Mimir v1's training cost, but the architecture and training paradigm are inherited from HRM-Text, suggesting broadly comparable economics for the compute-per-parameter portion of the run. What Mimir v1 adds is the specific permissible-data curation work — assembling, filtering, and licensing 161 datasets — which is a research investment not captured in a simple GPU-hour figure but is the element that makes the model institutionally deployable without legal review. That gap matters more now than it did two years ago. The AI training data copyright landscape as of August 2026 has produced significant legal benchmarks: in Thomson Reuters v. Ross Intelligence, a U.S. district court issued a landmark ruling finding that training a competing AI product on copyrighted legal content was not fair use. A court granted final approval in July 2026 to a $1.5 billion settlement between Anthropic and book authors, resolving claims tied specifically to training copies sourced from pirated material. Courts have drawn a legally material distinction between training on properly acquired or licensed data — which has generally survived fair-use analysis — and training on pirated or unlicensed content, which has not. Most high-performing small models currently available were fine-tuned on data distilled from GPT-4, Claude, or other frontier systems. That technique is fast and effective. It is also foreclosed for institutions that cannot accept proprietary license terms, and it has drawn regulatory scrutiny as AI companies increasingly prohibit the use of their outputs to train competing models. Mimir v1's permissible-data constraint is not merely an ethical posture — it is a structural commitment that makes the model usable in European research institutions, government agencies, and academic settings where data provenance must be auditable and legally defensible. Benchmark Results in Context The DFM team tested Mimir v1 across 20 benchmarks spanning English general knowledge, mathematics, code, and Danish language tasks. The paper reports that the model outperforms HRM-Text 1B — the closest comparable architecture — across the evaluation suite, and competes with Qwen 3.5 4B (released March 2, 2026) and Gemma 4 E2B (released April 2026, with approximately 2.3 billion effective parameters), models with roughly four times the parameter count. For reference on the parent architecture: HRM-Text 1B, trained on 40 billion unique tokens, achieved 60.7% on MMLU, 81.9% on ARC-Challenge, 82.2% on DROP, and 56.2% on MATH — competitive with 2- to 7-billion-parameter standard transformers at the time of its release. Mimir v1 claims to exceed those scores across the suite. On Danish-language benchmarks, the researchers report a new state of the art, which reflects both the project's mission and the relative scarcity of strong Danish-language training data in most competing models' pipelines. These results carry the standard caveat for a paper published a single day ago: they are self-reported by the authors, have not been independently replicated, and the 20-benchmark coverage, while broad, is not exhaustive. The HRM-Text paper itself found marginal benchmark contamination on the DROP dataset at the 13-gram level (though not at the 20-gram level), and its authors acknowledged that while contamination appeared not to drive results, the question deserves scrutiny for any follow-on models. A 2025 analysis of the original HRM architecture also noted that on some structured reasoning benchmarks, a plain transformer with comparable depth matched HRM's performance, raising the question of how much of the gain is specifically architectural versus simply the result of more total computation. Independent evaluation from the research community over the coming weeks will be the real test. Why Scale Assumptions Matter — and How Mimir Challenges One The AI field has operated, for several years, under a dominant training recipe: take a large decoder-only transformer, train it on internet-scale raw text, and improve it through instruction-tuning and reinforcement learning from human feedback. The Chinchilla scaling law, derived from extensive empirical work, prescribes jointly increasing parameters, data, and compute together — a prescription that has concentrated frontier-level capability in organizations with access to billion-dollar training infrastructure. HRM-Text's May 2026 result was already a significant challenge to that paradigm for reasoning-focused use cases: by co-designing architecture (recurrent depth over parameter width), objective (task-completion over next-token prediction on raw text), and data (40 billion curated instruction-response tokens), it entered the performance range of models trained with 100 to 432 times more compute. Mimir v1 now extends that challenge to the permissible-data domain: the same architectural efficiency that reduced compute requirements also, apparently, holds under a strict data-sourcing constraint. The broader implication is architectural. If HRM-based models continue to scale favorably and the Danish team's permissible-data results hold up under independent scrutiny, the compute and data divide that has structurally excluded most academic institutions from foundational AI research may be smaller than widely assumed. A public university with access to a 16-GPU node is not competing with OpenAI or Google on parameter count or training tokens. According to the Mimir v1 results, it may not need to be. What the HRM Architecture Cannot Yet Do The original HRM architecture, at its 27-million-parameter toy scale, was explicitly a task-specific reasoning engine — excellent at structured symbolic tasks (Sudoku, maze navigation, ARC-style problems) but not capable of open-ended text generation, code generation, or summarization. HRM-Text extended the architecture to language, but its creators openly acknowledged that coding performance lagged behind the model's reasoning and general-knowledge results — a direct consequence of training data composition rather than an architectural ceiling. Mimir v1 inherits this tradeoff. Its 161-dataset mixture is described as permissible post-training data across general instructions, rewritten knowledge, mathematics, symbolic tasks, reasoning traces, textbook exercises, and extracted web instructions — a richer mix than HRM-Text's initial corpus. Whether that mixture includes enough code-focused data to close the coding gap is a question the paper addresses across its benchmarks but that independent developers will need to verify in real-world deployment. The architecture's reliance on a PrefixLM attention pattern also means that inference must be set up correctly — with bidirectional attention over the prompt and causal attention over the response — or results will degrade noticeably. This is an engineering constraint rather than a conceptual limitation, but it adds deployment friction for developers using standard inference stacks that do not natively support split attention masking. Finally, the permissible-data claim itself merits independent audit. Dataset permissibility is a legal and institutional determination that depends on licensing interpretations that can be contested, and the 161-dataset mixture has not yet been externally reviewed. The DFM team's commitment to open, well-documented data practices — a hallmark of their prior work on Dynaword and EuroEval — suggests the claim is made in good faith, but "in good faith" and "legally defensible in all jurisdictions" are not identical. European researchers operating under GDPR with specific contractual data-use obligations should review the dataset documentation before relying on the permissibility designation. What Readers Can Do With This Now DFM Mimir v1 is available today at the Danish Foundation Models page on Hugging Face. The technical report arXiv:2608.13517 is freely available and runs 20 pages — readable in an afternoon for a practitioner evaluating whether the architecture and data sourcing approach fit their use case. For researchers constrained by data-ethics requirements, Mimir v1 is the most capable permissible-data model at the 1-billion-parameter scale currently documented. For developers evaluating small models for production deployment in institutions with legal data-provenance requirements, it is a new competitive baseline. For the broader research community, it is a result that will need replication and stress-testing — and the weights are available to begin that process today. Frequently Asked Questions What is the HRM architecture, and how is it different from a standard transformer language model? The Hierarchical Reasoning Model uses two coupled recurrent modules — a slow high-level planner and a fast low-level executor — that iterate over the same input embeddings in a nested loop during a single forward pass. Unlike transformers, which have a fixed computational depth set at build time, HRM can allocate more recurrent cycles to harder problems without adding parameters. It also trains on instruction-response pairs with a PrefixLM attention mask, concentrating every gradient update on generating correct answers rather than predicting raw text. The practical result is that HRM-Text 1B reached competitive benchmark performance on roughly 40 billion training tokens — a fraction of what comparable standard models consumed. What does "permissible post-training data" mean, and why does it matter legally? Permissible data, as the DFM team uses the term, refers to datasets sourced from materials that are in the public domain, explicitly licensed for AI training use, or otherwise legally cleared — as opposed to datasets scraped from the web under disputed copyright terms or distilled from proprietary models like GPT-4 or Claude. This distinction has become legally material: a U.S. court issued a landmark ruling in 2025 finding that training a competing AI on copyrighted content without authorization was not fair use, and a major copyright settlement in July 2026 resolved claims tied to training copies sourced from pirated material. Models distilled from proprietary teachers also face license-term restrictions in many institutional contexts. Mimir v1's permissible-data design makes it deployable in European academic and government settings that cannot assume fair-use protection or accept proprietary licensing terms — a practical institutional advantage that goes beyond ethics. What are the limitations of Mimir v1 that users should know before deploying it? Three practical limitations stand out. First, coding performance: like its predecessor HRM-Text 1B, Mimir v1's training data composition did not prioritize code, and code generation performance lags behind its reasoning and general-knowledge scores. Second, inference setup: the PrefixLM attention pattern requires that inference frameworks apply bidirectional attention to prompt tokens and causal attention to response tokens; standard inference stacks that do not support this will produce degraded results. Third, the benchmark results are self-reported from a paper published one day ago and have not yet been independently replicated — the community evaluation over the coming weeks will be the more reliable signal on where Mimir v1 actually stands against its claimed comparators. Can small open-source models trained on legal data seriously compete with commercial AI in 2026? Mimir v1 is a carefully documented data point suggesting the answer is "yes, for reasoning-focused use cases, more than most people assumed." The model was trained by a public university research team, on permissible data, at a training cost that is a tiny fraction of what commercial frontier models require, and it claims to match models four times its parameter count on 20 benchmarks. Independent replication will determine whether those claims hold broadly. What is clear is that the HRM architecture's core bet — that cycling fixed parameters intelligently yields more per compute dollar than simply adding parameters — has now been demonstrated in two papers from two different institutions. Whether it scales further is the open question. ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.