inference/research
verifiedmedium · 0.80updated 2026-10-18

Gemma 3n E2B config.json (unsloth mirror of google/gemma-3n-E2B-it)

repo

Fields

id
sources/hf-config-gemma3n-e2b-unsloth
Publisher
Hugging Face
Kind
repo
Accessed
2026-10-18
notes
google/gemma-3n-E2B-it config.json is license-gated (the official raw URL would not load). Read raw from the unsloth/gemma-3n-E2B-it mirror on 2026-10-18. text_config: num_hidden_layers 30, num_attention_heads 8, num_key_value_heads 2, head_dim 256, sliding_window 512, num_kv_shared_layers 10, max_position_embeddings 32768. layer_types is 24 sliding_attention plus 6 full_attention, repeating every 5 layers with the 5th global. The first 20 layers are the ones that allocate KV (16 sliding, 4 full); the last 10 reuse the KV of earlier layers. The file is a mirror, so confidence is below an official config.

Notes

google/gemma-3n-E2B-it config.json is license-gated (the official raw URL would not load). Read raw from the unsloth/gemma-3n-E2B-it mirror on 2026-10-18. text_config: num_hidden_layers 30, num_attention_heads 8, num_key_value_heads 2, head_dim 256, sliding_window 512, num_kv_shared_layers 10, max_position_embeddings 32768. layer_types is 24 sliding_attention plus 6 full_attention, repeating every 5 layers with the 5th global. The first 20 layers are the ones that allocate KV (16 sliding, 4 full); the last 10 reuse the KV of earlier layers. The file is a mirror, so confidence is below an official config.

Cross-references

This record references no other records.