Stops motion. Your device’s reduced-motion setting always takes priority.
A field guide to AI models
Find the model for the task.
Some models write. Others classify, transcribe or search through information. Explore Danish-developed models, Danish adaptations and selected international models, with model cards, licence terms and the available evidence of quality.
Around one billion parameters, trained from scratch using HRM-Text. Developed by SDU, Aarhus University, University of Copenhagen and the Alexandra Institute.
Possible public-sector use
Test local text processing, short drafts and bounded Danish-language questions.
Output
Text
Access & operation
Downloadable weights; requires the documented tokenizer and attention configuration.
Version, licence & sources+
Model ID
danish-foundation-models/DFM-Mimir
Released
· Official release date.
Licence / terms
Apache 2.0. Open weights; the licence alone does not establish the openness of the entire AI system.
Quality evidence
Researchers report results on 20 benchmarks. These are not demonstrated municipal production outcomes.
Worth checking4,096-token context. Not specifically safety-aligned; training languages are Danish and English.
Danish adaptation of Llama 2 7B from ITU. Included as a documented research reference, not the newest Danish model.
Possible public-sector use
Study Danish post-training and compare local experiments with newer models.
Output
Danish text
Access & operation
Downloadable weights; a separate 4-bit MLX version is available for Apple Silicon.
Version, licence & sources+
Model ID
NLPnorth/snakmodel-7b-instruct
Released
Not separately documented · Trained January–September 2024. Research paper published on 17 December 2024; a separate model release date is not established.
Licence / terms
Llama 2 Community License with specific conditions. Not an OSI open-source licence.
Quality evidence
Research paper covering eight Danish language and cultural tasks, compared with contemporary Llama 2 models.
Worth checkingOlder Llama 2 base and frequent factual errors according to the model card. Other languages are outside the intended scope.
Syv.ai fine-tuned Gemma 4 for Danish language tasks. The 12B variant supports classification, questions and Danish drafting.
Possible public-sector use
Test for sorting enquiries, reading comprehension and rewriting administrative text.
Output
Danish text
Access & operation
Downloadable weights for self-hosting.
Version, licence & sources+
Model ID
syvai/danskgpt-v4-12b
Released
Not separately documented · Exact release date is unconfirmed. Model card and access checked on 25 September 2026.
Licence / terms
CC BY 4.0 for the fine-tune; the Gemma 4 base model uses Apache 2.0. Syv’s card also refers generally to Gemma terms; clarify documentation when integrating.
Quality evidence
The card describes training; an independent Danish quality evaluation is unverified.
Worth checkingCan hallucinate and inherits base-model biases. Human review is necessary.
Syv.ai fine-tuned Gemma 4 for Danish language tasks. The 31B variant supports classification, questions and Danish drafting.
Possible public-sector use
Test for sorting enquiries, reading comprehension and rewriting administrative text.
Output
Danish text
Access & operation
Downloadable weights for self-hosting.
Version, licence & sources+
Model ID
syvai/danskgpt-v4-31b
Released
Not separately documented · Exact release date is unconfirmed. Model card and access checked on 25 September 2026.
Licence / terms
CC BY 4.0 for the fine-tune; the Gemma 4 base model uses Apache 2.0. Syv’s card also refers generally to Gemma terms; clarify documentation when integrating.
Quality evidence
The card describes training; an independent Danish quality evaluation is unverified.
Worth checkingCan hallucinate and inherits base-model biases. Human review is necessary.
Center for Humanities Computing · Aarhus Universitet
DaCy Large
Danish text analysis and named entity recognition / NER
Open weightsDanish-developed / Danish adaptation
A Danish model pipeline that locates names, organisations and places and analyses grammatical structure. It does not produce free-form chatbot answers.
Possible public-sector use
Could be tested for indexing documents and locating names for subsequent manual review.
Output
Text annotations, grammatical tags and relations between words.
Access & operation
Downloadable spaCy model for self-hosting; check its package version requirements.
Version, licence & sources+
Model ID
chcaa/da_dacy_large_trf · 0.2.0
Released
Not separately documented · The model card specifies version 0.2.0 but no separate launch date.
Licence / terms
Apache-2.0 for the model package and DaCy code. Open weights; no full assessment against the OSI AI definition has been made.
Quality evidence
The model card reports NER F1 of 87.38 on DaNE and describes Danish evaluation datasets. These are publisher-reported results.
Worth checkingNamed entity recognition is not complete anonymisation. The model can miss personal information; test on your own documents and add further review.
A Danish adaptation of mmBERT that identifies people, organisations and places in text. Training includes Danish, Norwegian and Swedish data.
Possible public-sector use
Could be tested for name-based document indexing and as one component in a personal-information review workflow.
Output
Entity annotations and scores for people, organisations and locations.
Access & operation
Weights can be downloaded from the publisher’s model repository for self-hosting.
Version, licence & sources+
Model ID
MediaCatch/mmBERT-base-scandi-ner-gold
Released
Not separately documented · The model card does not specify a separate release date.
Licence / terms
MIT according to the model card. Open weights; full open-source AI status has not been verified.
Quality evidence
The publisher reports aggregate F1 of 0.833 on a held-out test set. No separate Danish score is shown.
Worth checkingPrimarily trained on news and Wikipedia. It covers only three entity types and cannot anonymise text on its own. Test domain shifts and mixed-language text.
Scandinavian embedding model with 768 dimensions and up to 8,192 tokens. Based on scandmmBERT and DDSC data.
Possible public-sector use
Test search across municipal documents or grouping of enquiries.
Output
Numerical vectors, not text answers
Access & operation
Downloadable weights for local use with Sentence Transformers.
Version, licence & sources+
Model ID
emillykkejensen/mmBERTscandi-base-embedding
Released
Not separately documented · Exact launch date not established. Model repository last updated on 23 September 2025.
Licence / terms
The model card specifies Apache 2.0. The base model also uses Apache 2.0.
Quality evidence
Training is documented, but the developer described only internal tests when presenting the models. No verified municipal benefit or broad quality guarantee.
Worth checkingExperimental candidate. Test retrieval on your own domain language; a high similarity score does not prove a source is correct.
Scandinavian adaptation of Qwen’s 0.6-billion-parameter embedding model, trained on DDSC data.
Possible public-sector use
Test search across municipal documents or grouping of enquiries.
Output
Numerical vectors, not text answers
Access & operation
Downloadable weights for local use with Sentence Transformers.
Version, licence & sources+
Model ID
emillykkejensen/Qwen3-Embedding-Scandi-0.6B
Released
Not separately documented · Exact launch date not established. Model repository last updated on 14 September 2025.
Licence / terms
The model card specifies Apache 2.0. The base model also uses Apache 2.0.
Quality evidence
Training is documented, but the developer described only internal tests when presenting the models. No verified municipal benefit or broad quality guarantee.
Worth checkingExperimental candidate. Test retrieval on your own domain language; a high similarity score does not prove a source is correct.
Transcribes Danish speech. Trained on CoRal v3 conversational and read-aloud speech across dialects.
Possible public-sector use
Test recording transcription or subtitles, checking names and specialist terminology.
Output
Transcribed text
Access & operation
Downloadable weights; can run locally with Transformers.
Version, licence & sources+
Model ID
CoRal-project/roest-v3-wav2vec2-315m
Released
Not separately documented · Exact launch date not established. Model repository last updated on 19 June 2026; this is not a release date.
Licence / terms
Custom CoRal licence based on OpenRAIL-M, with use restrictions. Not unrestricted open source.
Quality evidence
The developer’s CoRal v3 test reports 13.7% character error rate on conversations. This is CER, not WER or a production guarantee.
Worth checkingThe licence restricts impersonation and identity inference, among other uses. Test noise, overlap and local dialects; transcription is not a finished meeting summary.
Transcribes Danish speech. Trained on CoRal v3 conversational and read-aloud speech across dialects.
Possible public-sector use
Test recording transcription or subtitles, checking names and specialist terminology.
Output
Transcribed text
Access & operation
Downloadable weights; can run locally with Transformers.
Version, licence & sources+
Model ID
CoRal-project/roest-v3-whisper-1.5b
Released
Not separately documented · Exact launch date not established. Model repository last updated on 19 June 2026; this is not a release date.
Licence / terms
Custom CoRal licence based on OpenRAIL-M, with use restrictions. Not unrestricted open source.
Quality evidence
The developer’s CoRal v3 test reports 11.6% character error rate on conversations. This is CER, not WER or a production guarantee.
Worth checkingThe licence restricts impersonation and identity inference, among other uses. Test noise, overlap and local dialects; transcription is not a finished meeting summary.
Jev reads text and answers bounded questions with choices, scores and probabilities. TypeSafe calls this a System One model; it does not generate free-form text.
Possible public-sector use
Could be tested for classifying enquiries by topic or assessing document relevance, with human review of uncertain results.
Output
Structured choices, scores and probabilities
Access & operation
TypeSafe-hosted API. Self-hosting or model-weight downloads are not documented.
Version, licence & sources+
Model ID
jev-1.13.0
Released
· This is the Jev family's public launch date. A separate release date for version 1.13.0 is not published in the reviewed sources.
Licence / terms
Proprietary API and customer terms. No public model-weight licence is documented.
Quality evidence
The description draws on the publisher’s documentation. This overview does not document an impact evaluation on Danish municipal tasks.
Worth checkingType-correct answers can still be wrong. English is the model's strongest language; Danish needs testing. TypeSafe documents weaknesses with arithmetic, dates and adversarial input, among others.
Language model with image understanding and reasoning
Closed model
A model for demanding writing, analysis and coding tasks. It can review documents and take part in longer workflows using tools.
Possible public-sector use
Could be tested for drafting professional notes, comparing documents and developing internal applications.
Output
Text and code from text and image inputs.
Access & operation
Hosted service and API, including through AWS, Google Cloud and Azure. Model weights are not available to download for self-hosting.
Version, licence & sources+
Model ID
claude-opus-5-5
Released
· Model release
Licence / terms
Closed model under commercial service terms; no public open-source or weights licence.
Quality evidence
The description draws on the publisher’s documentation. This overview does not document an impact evaluation on Danish municipal tasks.
Worth checkingCheck factual answers and evaluate Danish-language quality on your own tasks. Review data processing and tool permissions for the chosen service.
Language model for coding and tool-based workflows
Closed model
A model for complex coding tasks and workflows involving multiple tools. Sol can be used in Codex, but Codex is the agent environment, not the model itself.
Possible public-sector use
Could be tested for building and testing internal prototypes, processing documents and analysing material with approved tools.
Output
Text, code and structured responses from text and image inputs.
Access & operation
Hosted model in Codex, ChatGPT Work and the OpenAI API. Model weights are not available to download for self-hosting.
Version, licence & sources+
Model ID
gpt-6-sol
Released
· Model release
Licence / terms
Closed model under commercial service terms. Any open-source client code does not make the model open.
Quality evidence
The description draws on the publisher’s documentation. This overview does not document an impact evaluation on Danish municipal tasks.
Worth checkingReview code and results before use. Limit the agent's access and check data processing for the specific service and processing mode.
A model for summarisation, information extraction and focused coding tasks. It is a lower-cost option in the GPT-6 family for bounded tasks.
Possible public-sector use
Could be tested for structuring information from documents, producing short summaries and helping with repetitive administrative subtasks.
Output
Text, code and structured responses from text and image inputs.
Access & operation
Hosted model in Codex, ChatGPT Work and the OpenAI API. Model weights are not available to download for self-hosting.
Version, licence & sources+
Model ID
gpt-6-luna
Released
· Model release
Licence / terms
Closed model under commercial service terms; no public open-source or weights licence.
Quality evidence
The description draws on the publisher’s documentation. This overview does not document an impact evaluation on Danish municipal tasks.
Worth checkingSample-check extractions and summaries, and measure Danish-language quality. A lower price alone does not demonstrate that the model fits the task.
Marks when different speakers are active in audio, including overlapping speech. The model supports up to eight speakers in both streaming and recorded audio.
Possible public-sector use
Could be tested alongside an ASR model to attribute transcribed meeting or interview statements to different speakers.
Output
Generic speaker labels and timestamps
Access & operation
Model weights can be downloaded for self-hosting. Can be combined with a separate speech-to-text model.
Version, licence & sources+
Model ID
nvidia/Nemotron-3-Diarization
Released
· Release date confirmed in NVIDIA's model card and launch article.
Licence / terms
OpenMDW-1.1. Open weights with terms including preservation of the licence and notices when distributing.
Quality evidence
The description draws on the publisher’s documentation. This overview does not document an impact evaluation on Danish municipal tasks.
Worth checkingThis is diarization, not speech recognition: the model does not transcribe words or identify people by name. No separate Danish quality evaluation was found in the model card.
A specialist model with 600 million parameters that converts speech to text, both as a stream and from recordings. Danish belongs to the model's broad-coverage tier.
Possible public-sector use
Could be tested for dictation, captions and transcripts used as input for meeting minutes.
Output
Transcribed text with punctuation
Access & operation
Model weights can be downloaded and run on your own infrastructure with a compatible runtime.
Version, licence & sources+
Model ID
nvidia/nemotron-3.5-asr-streaming-0.6b
Released
· Release date stated in NVIDIA's model card.
Licence / terms
OpenMDW-1.1. The licence covers the supplied model materials; open weights alone does not establish fully open-source AI.
Quality evidence
The description draws on the publisher’s documentation. This overview does not document an impact evaluation on Danish municipal tasks.
Worth checkingTest Danish quality on your own terminology, accents and recordings. Of 40 language locales, 32 support transcription out of the box; eight need adaptation. The model does not write meeting minutes itself.