Avskrift CoreML model collection
Compiled model artifacts used by the Avskrift iOS app. These are model-weight adaptations and redistributed artifacts; runtime library licenses do not replace the model-weight licenses listed below. The collection does not imply endorsement by the upstream authors.
Model licenses
| Artifact | Source | Model-weight license | Runtime bytes including notices |
|---|---|---|---|
KBLab_kb-whisper-base |
KBLab/kb-whisper-base | apache-2.0 | 152,307,234 |
KBLab_kb-whisper-large_887MB |
KBLab/kb-whisper-large | apache-2.0 | 893,982,347 |
KBLab_kb-whisper-medium_649MB |
KBLab/kb-whisper-medium | apache-2.0 | 655,649,985 |
KBLab_kb-whisper-small |
KBLab/kb-whisper-small | apache-2.0 | 492,244,229 |
nemotron-3-diarization |
nvidia/Nemotron-3-Diarization | openmdw-1.1 | 398,228,671 |
opus-mt-sv-coreml |
Helsinki-NLP/opus-mt-tc-bible-big-deu_eng_fra_por_spa-gmq | apache-2.0 | 295,285,686 |
parakeet-tdt-0.6b-v3 |
FluidInference/parakeet-tdt-0.6b-v3-coreml | cc-by-4.0 | 483,274,654 |
pianissimo-sv-coreml |
KlangAI/pianissimo-sv | cc-by-4.0 | 632,188,981 |
Every artifact contains the applicable full license text and origin/adaptation notice under licenses/. KB-Whisper bundles also retain the OpenAI Whisper base-model MIT notice. provenance/ contains upstream model cards and recorded conversion settings. Original model-card results are upstream results, not new measurements of this release.
Formats and conversion
- Pianissimo: Klang's Swedish fine-tune of NVIDIA Parakeet v3, converted locally to the FluidAudio component contract. The compatible preprocessor is reused from FluidInference; the provenance manifest identifies its revision and conversion details.
- KB-Whisper base/small: CoreML float16 conversion with whisperkittools. Medium uses 6-bit palettization and large uses 4-bit palettization; embeddings remain float16. The folders retain their existing app identifiers.
- Parakeet v3: compiled components from FluidInference revision
aed02740059203c4a87495924f685de3722ae9ce, without additional re-conversion. - Nemotron 3 Diarization: existing FluidInference monolithic v2
fastandofflinefiles, silence embedding and GA 2026-09-23 weights marker. This is speaker diarization, not the Nemotron 3.5 ASR model. The current upstream CoreML card notes an offline ANE compiler limitation; the model must be tested with the app's actual runtime configuration.
Installation contract
Each model directory has files.json (relative path, byte count and SHA-256 per runtime file) and version.json (upstream version and conversion). Avskrift pins downloads to a full Hugging Face commit and verifies the raw manifest hashes recorded in its trusted catalog, then each file's SHA-256 before activation. CoreML preparation still occurs separately on each device.
The mutable catalog.json is published only after the artifact commit is known. It refers to that immutable full commit, and existing downloads continue using their original pinned revision.
This collection contains no original audio, user transcripts, credentials or app account information.