hviske-v6
Copyright (c) 2026 Syv.ai

The Python modelling, configuration and inference code in this repository is a
DERIVATIVE WORK of saga-2 by Capacit A/S:

    saga-2
    Copyright (c) 2026 Capacit A/S
    https://huggingface.co/capacit-ai/saga-2-m

It was derived from a revision published under the Apache License, Version 2.0
(the licence saga-2 carried up to and including commit
f1c014f383d59be1302c4eac350f6335a30b8f68; saga-2 moved to CC BY-NC 4.0 for
revisions published from 2026-09-24). A copy of the Apache License is in
LICENSES/Apache-2.0.txt.

Files derived from saga-2, and what they were derived from:

    modeling_whisper_qwen.py      saga2/model.py, saga2/encoder_pos.py
    configuration_whisper_qwen.py saga2/model.py (WQwenConfig)
    processing_whisper_qwen.py    saga2/gpu_mel.py, saga2/decode.py
    vllm/hviske_vllm/encoder.py   saga2/encoder_pos.py
    vllm/hviske_vllm/audio.py     saga2/gpu_mel.py, saga2/decode.py

All of these files have been modified. The changes are summarised at the top of
each file, as Apache License 2.0 section 4(b) requires.

The model WEIGHTS in this repository are also a derivative work of saga-2. The
encoder was initialised from the encoder tensors of the first release of
capacit-ai/saga-2-m -- published under the Apache License, Version 2.0 -- and was
then trained further on the corpus described in the model card. The decoder was
initialised from Qwen/Qwen3-0.6B-Base and the projector was randomly initialised;
those two parts are not derived from saga-2.

Modification, per Apache License 2.0 section 4(b): the saga-2 encoder weights have
been changed by further training. The RoPE adaptation of the Whisper encoder and
the Danish training preceding this repository are Capacit A/S's work.

This repository's own weights are licensed CC BY-NC 4.0; see the model card.

Third-party components
----------------------
Encoder: capacit-ai/saga-2-m, Copyright (c) 2026 Capacit A/S, Apache License 2.0
  (the licence carried by the revision this was derived from). See
  LICENSES/Apache-2.0.txt. saga-2-m is in turn derived from
  openai/whisper-large-v3-turbo, Copyright (c) 2022 OpenAI, MIT License. See
  LICENSES/MIT-Whisper.txt.
Decoder: Qwen/Qwen3-0.6B-Base, Copyright 2024 Alibaba Cloud,
  Apache License 2.0. See LICENSES/Apache-2.0.txt.

Training corpora carry their own licenses.