kyrr-models

Models that kyrr downloads for dictation. The same files are attached to the models-* releases of heast13/kyrr-assets, a second mirror. Each folder has its own license, see LICENSE.md.

de-primeline-u8

primeline/parakeet-primeline by primeLine AI, a German fine-tune of NVIDIA parakeet-tdt-0.6b-v3, both licensed under CC-BY-4.0.

Changes made for kyrr, with tools/export-primeline:

• exported from the NeMo checkpoint to ONNX as encoder, decoder and joiner for sherpa-onnx
• quantized to int8 with ONNX Runtime's dynamic quantization: the encoder's MatMul/Gemm weights per channel (its convolutions stay fp32), decoder and joiner per tensor
• the int8 weights of encoder and joiner stored as uint8 (+128 on weights and zero points, exact), so the model also works on x86 CPUs without VNNI; nothing else changed

File	Size	SHA256
de-primeline-u8/encoder.int8.onnx	881,596,736	49596a2c9c20a224fd087b2c1dc8cdc5dfe3fb8d3ddd4eefb070d81e492c93d8
de-primeline-u8/decoder.int8.onnx	11,845,274	6b32c9339900086afd3d282eb40688ae19f3d106281e7731ee27ead1add8b29d
de-primeline-u8/joiner.int8.onnx	6,355,277	c84fcf848108531a7309ef65308dc5dc610091ce0fbde134f76bb0f6c42ca530
de-primeline-u8/tokens.txt	93,939	d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d

en-unified

nvidia/parakeet-unified-en-0.6b, Licensed by NVIDIA Corporation under the NVIDIA Open Model License; a copy of the license is in en-unified/NVIDIA-Open-Model-License.pdf (version of October 24, 2025, from nvidia.com).

ONNX int8 conversion for sherpa-onnx by the sherpa-onnx project, from csukuangfj2/sherpa-onnx-nemo-parakeet-unified-en-0.6b-int8-non-streaming at revision 8c3a10fb13408c7a7054f6898958bf1c64a8d6c7. Mirrored unchanged, so kyrr's downloads do not depend on that repository.

File	Size	SHA256
en-unified/encoder.int8.onnx	654,040,552	6716910b7a0833997fec7a410494c995d70124001a0e9b66d6370d6aced577e0
en-unified/decoder.int8.onnx	7,257,753	a5e223392c90e75f8144cdb5eb95af7625db389e39edef2bd1a9c872b3298fe6
en-unified/joiner.int8.onnx	1,735,860	869f43f7d24595c55581ad3bf249a935fb8a71389fbdaa7504b9f46f93140f8a
en-unified/tokens.txt	8,952	dc0b4584ab2e4ddbf888425c076c61b736e7356a015250db7d307e6f1a8188ff

de-primeline (superseded)

The first upload, used by kyrr 0.2.0: the same model with signed int8 weights in encoder and joiner. ONNX Runtime gets these wrong on x86 CPUs without VNNI (Intel before Ice Lake/Alder Lake, AMD before Zen 4). Kept so the pinned links of older kyrr versions keep working.