gemma-4-E2B-it-gf17

Gemma-4 E2B IT — GF(17) Bake for Adam

This is a lossless GF(17) re-encoding of Google's Gemma-4 E2B Instruct weights, baked for use with Adam — a texture-native local AI assistant.

What is this?

The original Apache 2.0 Gemma 4 E2B IT weights are re-encoded as 4-tier GF(17) RGBA pixel atlas (PTEX) pages. Round-trip cosine similarity vs. source fp16 = 1.0 (bit-exact reversible — no fine-tuning, no lossy compression). This format enables:

• ~4 GB VRAM inference (vs ~6 GB for the safetensors equivalent)
• TMU lookup via GPU texture units (vs GEMM matmul)
• Progressive tier-streaming from NVMe → VRAM

Usage

Auto-downloaded by Adam's install script. No manual setup needed:

git clone https://github.com/Amnibro/Amni-Ai
cd Amni-Ai
python install.py

Adam will detect the bake, download it (~20 GB), and start the server at http://127.0.0.1:8002/.

Manual download

from huggingface_hub import snapshot_download
snapshot_download(repo_id='amnibro/gemma-4-E2B-it-gf17', local_dir='./bakes/gemma4_e2b_it_gf17')

License & Modification Notice (Apache 2.0 § 4)

The weights distributed in this repository are a modification of Google's Gemma 4 E2B IT model. Per Apache License 2.0 § 4:

• § 4(a): A copy of the Apache 2.0 License is included in this repository as LICENSE.
• § 4(b) (modification notice): The Gemma 4 E2B IT fp16 weights have been losslessly re-encoded into a 4-tier GF(17) RGBA pixel atlas (the "Reffelt decomposition"). Round-trip cosine similarity against the original fp16 weights is 1.0 (bit-exact reversible). No fine-tuning, distillation, or lossy compression has been applied — this is purely a storage-format conversion. See NOTICE for details.
• § 4(c): All copyright, patent, trademark, and attribution notices from the upstream Gemma 4 E2B IT distribution are preserved in NOTICE.
• § 4(d): The NOTICE text file is included alongside the weights.

Upstream source: https://huggingface.co/google/gemma-4-e2b-it Upstream license documentation: https://ai.google.dev/gemma/docs/gemma_4_license

Adam's surrounding orchestration code (the Amni-Ai repository) is CC BY-NC 4.0 — that license applies only to Amnibro's original code, not to these weights, which remain under Apache 2.0.

Trademark

Gemma is a trademark of Google LLC. This bake is not endorsed by, sponsored by, or affiliated with Google. Reference to Gemma is purely attributive in compliance with Apache 2.0 § 4(c).