Evidence-backed European AI register
KBLab — National Library of Sweden
Reviewed identity, roles, model listings, pricing and evidence for KBLab — National Library of Sweden.
Dataset compiled 2026-08-27 · evidence over inference
Reviewed identity
- Headquarters scope
- EU headquarters
- Headquarters record
- Stockholm
- Legal entity
- data lab within Kungliga biblioteket (the National Library of Sweden), a Swedish public authority
- Ownership
- state-owned / public research
- Founded
- 2019
- Confidence
- high
- Last verified
- 2026-08-22
Evidence boundaries
Organisation listings identify a reviewed canonical record. Model listings show an association in the source register; they do not by themselves establish model creation, current API availability, hosting location or infrastructure ownership.
Model listings
- KB-Whisper (tiny / base / small / medium / large)
- Type
- asr
- Parameters
- 57.7M, 99.1M, 0.3B, 0.8B, 2B
- Context
- unknown
- Languages
- Swedish
- Licence
- unknown
- Release
- 2025-03
- Type
- text
- Parameters
- ~110M
- Context
- 512 tokens
- Languages
- Swedish
- Licence
- unknown
- Release
- 2020
- bert-base-swedish-cased-new
- Type
- text
- Parameters
- 110M
- Context
- 512 tokens
- Languages
- Swedish
- Licence
- unknown
- Release
- unknown
- megatron-bert-large-swedish-cased-165k
- Type
- text
- Parameters
- 0.4B
- Context
- 512 tokens
- Languages
- Swedish
- Licence
- unknown
- Release
- unknown
- sentence-bert-swedish-cased
- Type
- embedding
- Parameters
- 0.1B
- Context
- unknown
- Languages
- Swedish
- Licence
- unknown
- Release
- unknown
- wav2vec2-large-voxrex-swedish (VoxRex)
- Type
- asr
- Parameters
- 0.3B
- Context
- unknown
- Languages
- Swedish
- Licence
- unknown
- Release
- unknown
- bart-base-swedish-cased
- Type
- text
- Parameters
- 0.1B
- Context
- unknown
- Languages
- Swedish
- Licence
- unknown
- Release
- unknown
- Swedish OCR correction
- Type
- text
- Parameters
- 0.3B
- Context
- unknown
- Languages
- Swedish
- Licence
- unknown
- Release
- unknown
- bert-base-swedish-cased-ner
- Type
- text
- Parameters
- 110M
- Context
- 512 tokens
- Languages
- Swedish
- Licence
- unknown
- Release
- unknown
Profile note
The strongest argument in Europe for national libraries as model labs: KB-Whisper, trained on 50,000+ hours from the Swedish legal-deposit broadcast archive, cuts WER 47% versus OpenAI's whisper-large-v3 on Swedish. KBLab produces no chat LLM — its output is encoders, embeddings, ASR and OCR correction, all free. Licences are not consistently surfaced on the org page and must be read per model card. No commercial API.