Gemma 4 E4B Model Verified and Ready | Generated by AI
It works! Here’s the complete verification:
Model loaded and generated text: “I am a language model, trained by Google.”
Architecture: Gemma 4 E4B (7.52B params, Q3_K quant, 3.76 GiB file)
Gemma 4 E4B key specs from the GGUF metadata:
- 42 layers, 2560 embed dim, 10240 FF dim
- 8 attention heads, 2 KV heads (GQA=4)
- 131K context length (both global + sliding window attention at 512)
- BPE tokenizer, 262k vocab
- Shared KV layers (18 shared across layers — Gemma 4’s layer-tied KV design)
- Layer Input weights («per_layer_token_embd», per-layer input gating/projection — Gemma 4’s “Gated Delta Net” architecture)
Performance on RTX 4070 12GB:
- Prompt eval: 291 t/s
- Generation: 91 t/s
- VRAM used: 2343 MiB model + 14 MiB KV + overhead → ~3.2 GiB total, plenty of room
The file is at /mnt/data/models/gemma-4-E4B-it-Q3_K_M.gguf and ready to use. Want me to start the llama-server so you can chat with it or configure it in Hermes?
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋