Gemma 4 Model Setup and llama.cpp Update | Generated by AI

Home 2026.07

Here’s the full session summary.


Session: Gemma 4 E4B — Download, verify, update llama.cpp

1. Downloaded model

2. Model architecture (from GGUF metadata)

Property Value
Architecture Gemma 4 (gemma4)
Params 7.52 B (labeled “E4B”)
Layers 42
Embed dim 2560
FF dim 10240
Attention heads 8 (GQA = 4, 2 KV heads)
Context length 131,072
Sliding window 512 (hybrid — 35 SWA layers + 7 global layers)
RoPE base 1,000,000 (global), 10,000 (SWA)
Logit softcap 30.0
Shared KV layers 18 (layer-tied KV design)
Per-layer input embeds Yes (Gated Delta Net)
Vocab BPE, 262,144 tokens
License Apache 2.0

3. Initial inference test (old build b8)

4. Updated llama.cpp

5. Build

6. Final verification

Current state


Back Donate