dense, 60 layers, encoder-free 12B variant; 26B-A4B MoE 3.8B active; E2B/E4B small
Gemma 4 is Google's newest generation, spanning E2B, E4B, 12B, 26B-A4B MoE and 31B dense sizes with an encoder-free 12B variant, released in April 2026 under the Apache-2.0 license. Its quantization story centers on the official QAT Q4_0 checkpoint (31B = 19.2 GB) and the nvidia NVFP4 build, which stays within half a point of its baseline on six benchmarks.
Gemma 4 is a 30.7B dense Apache-2.0 model from Google with 256K (12B) / 128K (E2B/E4B) context. Quantized checkpoints lose 0.22–0.45 points on the 5 measured per-quant benchmarks.
| Checkpoint | Family | Size | Source |
|---|---|---|---|
| Google Gemma-4-31B-it QAT Q4_0 GGUF Official QAT Q4_0 (31B) | GGUF | 19.2 GB | source ↗ |
| nvidia Gemma-4-31B-IT-NVFP4 21B params (NVFP4); size not captured | NVFP4 | — | source ↗ |
| unsloth gemma-4-12b-it-GGUF Q4_K_M 12B | GGUF | 7.12 GB | source ↗ |
| unsloth gemma-4-12b-it-GGUF Q5_K_M 12B | GGUF | 8.41 GB | source ↗ |
| unsloth gemma-4-12b-it-GGUF Q6_K 12B | GGUF | 9.79 GB | source ↗ |
| unsloth gemma-4-12b-it-GGUF Q8_0 12B | GGUF | 12.7 GB | source ↗ |
| unsloth gemma-4-12b-it-GGUF BF16 12B | BF16 | 23.8 GB | source ↗ |
| cyankiwi gemma-4-26B-A4B-it-AWQ-4bit 26B-A4B | AWQ | 17.19 GB | source ↗ |
| turboderp Gemma 4 EXL3 EXL3; size not captured | EXL3 | — | source ↗ |
baseline 85.8
baseline 87.92
baseline 85.25
baseline 82.49
baseline 33.61
baseline 27.08