JoshHughes/gemma4-e2b-q4-lma

Gemma 4 E2B-it, balanced 4/8-bit, one file (LMA)

Google's Gemma 4 E2B-it in one 3.3 GB file, for on-device web pages that use Local Model Access (LMA) - created by artist Josh Hughes. It answers text and pictures.

Download it once into your device's files (on iPhone: Files › Downloads, with Safari's own downloader). Every LMA page that approves it then picks the same copy: nothing is uploaded and nothing is copied, and each page checks the file once by its sha256 and remembers.

File

File	gemma4-e2b-q4-lma.safetensors
Size	3,263,162,540 bytes
sha256	c29e8d9e9c96e0b95104866adacf11bb477dbcd91fa83e9bc8f11eb510a137df

What is in it

• The source pack, byte for byte. It is JoshHughes/p0pc0rn-assistant-gemma4-e2b-q4 at revision 141f0e8665691d74f824c4469725cf19a2ded9e8 (manifest sha256 e34bc23f7aa8c17244b53068e402a55fbc0ec68fb989b42e2615c7439d792178). Its format is pocketflux.pack/1, the 93 files of the source pack, in one safetensors file:

  • each file is a 1-D U8 tensor named pack: + its shard name, with its bytes unchanged;
  • the pack's manifest is in the file's metadata, under pocketflux.pack;
  • the data starts on a 4 KiB boundary.

  The layout is specified in LMA's SPEC.md §2.2. LMA's tools/pack_to_safetensors.py made it.

• Quantisation (the source pack's, unchanged):

  • the language model's MLPs are GPTQ-fitted 4-bit (groups of 32);
  • attention, the gates, the per-layer projections and the MLPs of layers 4, 6, 12 and 13 are 8-bit;
  • the vision tower is 8-bit;
  • the embedding and the per-layer embeddings are 4-bit;
  • the norms are f32.

  The encodings are described in LMA's docs/QUANT-q4g32a.md.

• The tokenizer is Gemma 4's, in a compact form. The chat template is verbatim.

It is not a Transformers checkpoint. LMA's engine (lma/engine/) runs it on WebGPU, reading the file a slice at a time from the visitor's own copy. On a desktop GPU, this file and the source pack give bit-identical logits.

Licence and credit

• Gemma 4 E2B-it is by Google, released under the Apache License 2.0 (LICENSE: the Apache-2.0 text; Google's licence page: https://ai.google.dev/gemma/docs/gemma_4_license), not the Gemma Terms of Use. "Gemma" is a Google trademark.
• The weights here were converted and quantised from Google's original by the p0pc0rn project (Josh Hughes), and repackaged as one file for Local Model Access. NOTICE.md says what changed.
• Original model: google/gemma-4-E2B-it at revision 3e22461f65e89153144f8adb70e3b8c2cc9845a7 (model.safetensors sha256 2db5482b20d746879bb3ef79b5203e9075a2e2b98f54ec7c2f281c1477ddc550).