| # Licenses and attribution | |
| This repository distributes an independently modified NVIDIA Mamba-2 8B model and its matching software. Different components retain different license scopes; the Hugging Face `license: other` field describes this mixed distribution. | |
| - **Upstream base model and tokenizer:** NVIDIA declares Apache-2.0. The model weights have been quantized and readapted, and a separately trained recall adapter has been added. This is not an official NVIDIA release or endorsement. | |
| - **Project software, native adapter implementation, and QuIP#-derived E8/LDLQ implementation:** GPL-3.0. Matching source is included in `release/source.zip`; the software license also applies to this repository's new wrapper documentation. Software-derived resources retain their provenance. | |
| - **Model versus software:** use of a GPL quantizer does not by itself assign a new license to all input/output weights. The retained component notices govern their respective scopes; this page does not replace them or assign an additional standalone license to the adapter artifact. | |
| The original release's notices and license texts are included without modification: | |
| - [Component scope and provenance](release/licenses/THIRD_PARTY.md) | |
| - [NVIDIA model card and attribution](release/licenses/NVIDIA_MAMBA2_MODEL_CARD.md) | |
| - [Apache-2.0](release/licenses/APACHE-2.0.txt) | |
| - [Mamba runtime Apache-2.0 notice](release/licenses/MAMBA-APACHE-2.0.txt) | |
| - [Project GPL-3.0](release/licenses/PROJECT-LICENSE) | |
| - [QuIP# GPL-3.0](release/licenses/QUIP-SHARP-GPL-3.0.txt) | |
| The [source repository](https://github.com/EndlessChasing/mamba2-8b-e8w5) records the modifications, authorship acknowledgements and evaluation-data provenance. Private reference artifacts and WikiText passages are not part of this distribution. | |