Nemosine AI Model Notices
=========================

The following notices apply to optional AI model files downloaded by
Nemosine. The models run locally on the user's device.


Gemma 4
-------

Gemma 4 is developed by Google DeepMind. Nemosine uses instruction-tuned
Gemma 4 E2B, E4B, 12B, 26B A4B, and 31B models for optional on-device chat.

License: Apache License 2.0

Nemosine downloads LiteRT-LM conversions published by the LiteRT Community.

Original models:
https://huggingface.co/google/gemma-4-E2B-it
https://huggingface.co/google/gemma-4-E4B-it
https://huggingface.co/google/gemma-4-12B-it
https://huggingface.co/google/gemma-4-26B-A4B-it
https://huggingface.co/google/gemma-4-31B-it

LiteRT-LM conversions:
https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm
https://huggingface.co/litert-community/gemma-4-E4B-it-litert-lm
https://huggingface.co/litert-community/gemma-4-12B-it-litert-lm
https://huggingface.co/litert-community/gemma-4-26B-A4B-it-litert-lm
https://huggingface.co/litert-community/gemma-4-31B-it-litert-lm


NeoN and NeoN Plus (MiniCPM5)
-----------------------------

NeoN is powered by MiniCPM5 1B, and NeoN Plus is powered by MiniCPM5 2B.
MiniCPM5 is developed by OpenBMB. Nemosine uses these models for optional
on-device chat. NeoN and NeoN Plus are the names Nemosine shows for them;
they are not OpenBMB products.

License: Apache License 2.0

Original models:
https://huggingface.co/openbmb/MiniCPM5-1B
https://huggingface.co/openbmb/MiniCPM5-2B

Nemosine downloads the Q8_0 GGUF files published by OpenBMB, unmodified:
https://huggingface.co/openbmb/MiniCPM5-1B-GGUF
https://huggingface.co/openbmb/MiniCPM5-2B-GGUF


K2 Horizon 3.7B
---------------

K2 Horizon 3.7B is developed by the Institute of Foundation Models (IFM) at
MBZUAI. Nemosine uses it for optional on-device chat on Mac.

License: Apache License 2.0

Original model:
https://huggingface.co/IFM/K2-Horizon-3.7B

Nemosine downloads the Q8_0 GGUF file published by IFM, unmodified:
https://huggingface.co/IFM/K2-Horizon-3.7B-GGUF


Granite Embedding 311M Multilingual R2
--------------------------------------

Granite Embedding 311M Multilingual R2 is developed by the Granite Embedding
Team at IBM. Nemosine uses it for optional on-device search and retrieval.

License: Apache License 2.0

Original model:
https://huggingface.co/ibm-granite/granite-embedding-311m-multilingual-r2

Nemosine uses the LiteRT conversion published by the LiteRT Community:
https://huggingface.co/litert-community/granite-embedding-311m-multilingual-r2

Modification notice from the LiteRT distribution: the files were converted
from the IBM model. The weights were exported to LiteRT and quantized using
int8 dynamic-range quantization. CLS pooling and L2 normalization were folded
into the graph. No fine-tuning or weight modification beyond quantization was
performed.


EmbeddingGemma 2
----------------

EmbeddingGemma 2 is developed by Google DeepMind. Nemosine uses its text
model for on-device search experiments.

License: Apache License 2.0

Original model:
https://huggingface.co/google/embeddinggemma-2

Nemosine downloads the LiteRT-LM conversion published by the LiteRT
Community, unmodified:
https://huggingface.co/litert-community/embeddinggemma-2-text-270m-litert-lm

Nemosine also uses its own Core ML conversion of the text model, made from
the original weights. Modification notice: the text model's weights were
converted from bfloat16 to float16 and re-expressed for Apple's Neural Engine
(channels-first layout, 1x1 convolutions in place of linear layers, per-head
attention, fixed sequence lengths). RMSNorm rescales each token before
squaring, and mean pooling happens before the final projection; both are
algebraically identical to the original. The embedding scales were folded
into the weights, and mean pooling and L2 normalization were added to the
graph. The vision and audio encoders are not included. No fine-tuning,
quantization or pruning was performed.

Use of EmbeddingGemma 2 is subject to the Gemma Prohibited Use Policy:
https://ai.google.dev/gemma/prohibited_use_policy


multilingual-e5-small
---------------------

multilingual-e5-small is developed by Liang Wang, Nan Yang, Xiaolong Huang,
Linjun Yang, Rangan Majumder and Furu Wei ("Multilingual E5 Text Embeddings:
A Technical Report", arXiv:2402.05672). Nemosine uses it for on-device search
experiments.

License: MIT License

Original model:
https://huggingface.co/intfloat/multilingual-e5-small

Nemosine uses its own Core ML conversion of the model, together with the
model's tokenizer files. Modification notice: the weights were re-expressed
for Apple's Neural Engine (channels-first layout, 1x1 convolutions in place
of linear layers, per-head attention, a fixed 256-token sequence length) and
stored as float16, with the token embedding table quantized to int8
(per-token scales). Mean pooling and L2 normalization were added to the
graph. No fine-tuning or other weight modification was performed.

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.


The complete Apache License 2.0 text is included with Nemosine and is
available from the AI model notices section in Legal Information.

The names IBM, Granite, Google DeepMind, Gemma, OpenBMB, MiniCPM, IFM,
MBZUAI, K2 Horizon, LiteRT, and LiteRT Community are used only to identify
the origins of the distributed works. Their use does not imply endorsement
of Nemosine.
