Gemma4-E2B-Dflash2-Android

Model weights: Apache License 2.0. Original Gemma 4 E2B instruction-tuned model
Copyright Google; https://huggingface.co/google/gemma-4-E2B-it
Original weights are redistributed unchanged. The unmodified Q4 GGUF is from
https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF (Apache-2.0).
The compatible mmproj is an external download from that upstream repository;
it is not redistributed in this package. Google provides the original weights.

Modified artifacts: mixed-precision quantization, QNN compilation, intermediate
feature outputs, portable vision-embedding operator alias, CPU drafter training
and packaging. These are independent preparations, not official Google,
Qualcomm, Unsloth or Z Lab releases. Detailed source revisions and changes are
recorded in provenance.json and quantization.json.

DFlash2-derived grouped dynamic-convolution/selection components are based on
https://github.com/z-lab/dflash at 07ebd93db9f472af339b644bb70221ad8428328a,
Copyright (c) 2026 Z Lab, MIT. The locally trained draft weights are distinct
from upstream checkpoints; algorithm lineage does not imply endorsement.
Full MIT notice is retained in licenses/DFlash-MIT.txt.

RaInfer engine source: https://github.com/LEPTONNW/RaInfer, MIT.
llama.cpp/mtmd source: https://github.com/ggml-org/llama.cpp, MIT.
ExecuTorch source: https://github.com/pytorch/executorch, BSD-3-Clause.
Separate license texts are preserved in licenses/ for attribution. Engine
sources, applications, binaries and proprietary SDK/runtime files are not
distributed in this model package. Their respective licenses remain separate.