License & Attribution
This mirror redistributes Google DeepMind's Gemma 4 weights under
Apache 2.0, the original license. All credit for the model goes to
Google DeepMind.
This is NOT a derivative work — every file in this repo is byte-identical
to the canonical source as of the mirror date below. Use the canonical Google
repo when available; this mirror exists as a redundancy.
Verification
All small files (config, tokenizer, README, etc.) were SHA-256 verified
bit-identical to the canonical source at the time of upload. The 10.2 GB
model.safetensors is content-addressed by HF LFS and the ETag matched.
Table with columns: Field, Value| Field | Value |
|---|
| Canonical | google/gemma-4-E2B-it |
| Mirror | xaitalk/gemma-4-E2B-it-mirror |
| Mirror date | 2026-05-28 |
| Total size | ~10.3 GB |
| Files mirrored | 9 (all) |
How to use
This mirror is a drop-in replacement for the canonical repo:
from transformers import AutoModelForMultimodalLM, AutoProcessor
model = AutoModelForMultimodalLM.from_pretrained("google/gemma-4-E2B-it")
model = AutoModelForMultimodalLM.from_pretrained("xaitalk/gemma-4-E2B-it-mirror")
Inside the xaitalk framework, xaitalk.hub.ensure_model("gemma-4-e2b-it")
tries the canonical first and falls back to this mirror automatically.
Why xaitalk mirrors this model
xaitalk is a cross-framework XAI library
that supports Gemma 4 in PyTorch, TensorFlow, and JAX with bit-equivalent
attribution methods. Mirroring the canonical weights guarantees long-term
reproducibility of the cross-framework benchmarks shipped with the library.
Cross-framework results for Gemma 4 E2B (8 methods × 2 modalities,
fp32, A100 80GB, 2026-05-28):
Table with columns: Modality, seq, methods, min r| Modality | seq | methods | min r |
|---|
| image+text | 285 | 8/8 PASS | r ≥ 0.9999998 |
| audio+text | 41 | 8/8 PASS | r ≥ 0.9999999 |
Mirror policy
Per the xaitalk hub-mirror policy, third-party artifacts ≤ 2.5 GB are
mirrored by default. Gemma 4 E2B at ~10 GB is an explicit exception
because it's the flagship "any-to-any" thinking model in xaitalk's
coverage matrix, and license compliance (Apache 2.0) permits
redistribution with attribution.
— xaitalk