QmogAI/Gemma4-E2B-int4.mog

Gemma4-E2B-int4

Gemma 4 E2B (instruction-tuned, QAT) in .mog format. One file: config, tokenizer, and text weights.

Exported from google/gemma-4-E2B-it-qat-q4_0-unquantized with qpack. Vision and audio towers are omitted; this pack is text-only.

Base	google/gemma-4-E2B-it-qat-q4_0-unquantized
Format	MOG v2, int4
Size	~2.5 GB
Weights	Group-wise symmetric int4 (group size 64) on attention/MLP projections and embedding tables; norms stay f16
Architecture	gemma4 (hybrid sliding/full attention, dual RoPE, KV sharing)
Context	config max position embeddings: 131,072

Download

hf download QmogAI/Gemma4-E2B-int4.mog gemma4-E2B-int4.mog --local-dir .

Export your own

pip install -r requirements.txt

python export.py \
  --model_dir /path/to/gemma-4-E2B-it-qat-q4_0-unquantized \
  --quant int4 \
  --out gemma4-E2B-int4.mog

License

Apache 2.0, following the Gemma 4 license of the base checkpoint.