p0pc0rn pack: Assistant (Gemma 4 E2B-it, balanced 4/8-bit)
Converted and quantized by the p0pc0rn project from Google's Gemma 4 E2B-it. Modified by the p0pc0rn project: converted and quantized from the original weights.
- Language model: GPTQ 4-bit (groups of 32) MLPs, with 8-bit attention, gates and the MLPs of layers 4, 6, 12 and 13; vision tower 8-bit.
- Tokenizer: Gemma 4's, stored in a compact form; the chat template verbatim.
- Source:
google/gemma-4-E2B-it@3e22461f (Apache-2.0, not the Gemma Terms of Use).
Size: 3.26 GB in 93 shards. Licence: Apache-2.0 (LICENSE). Notices: NOTICE.md.
About p0pc0rn
p0pc0rn is an on-device app by Joshua Hughes: pictures, short videos, songs and music videos, made by WebGPU on the phone. Where a licence asks for it (LTX-2, FLUX Non-Commercial), the app shows the licence and asks you to accept it before the first byte downloads; every licence is listed in the app under Settings › About › Licences and notices. The app checks every 16 MiB chunk against manifest.json (sha256) as it lands.
Format. pocketflux.pack/1: manifest.json lists every shard, every tensor (offset, shape, encoding) and the sha256 of every chunk; the .pfw shards are raw tensor data (group-wise 4-bit / 8-bit or 16-bit floats). Only p0pc0rn's WebGPU engine reads them. This is not a Transformers or Diffusers checkpoint.
Contact: X @p0pc0rnapp (official) or @joshhughes___ (Josh).