Gemma 4 · released 2026-06 · 12.0B parameters
'Unified' model (Gemma4UnifiedForConditionalGeneration): image/audio enter via linear projections, no separate encoders. Global layers: 1 KV head x 512, K=V.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.56 GB | 0.28 GB |
| 8,192 tokens | 0.81 GB | 0.41 GB |
| 32,768 tokens | 2.31 GB | 1.16 GB |
| 131,072 tokens | 8.31 GB | 4.16 GB |
| 262,144 tokens | 16.3 GB | 8.16 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 22.3 GB |
| FP8 (W8A8) | 8 | 12.1 GB |
| INT8 (W8A8 / W8A16) | 8 | 12.1 GB |
| NVFP4 | 4.5 | 7.61 GB |
| INT4 AWQ / GPTQ | 4.25 | 7.29 GB |
| GGUF Q8_0 | 8.5 | 11.8 GB |
| GGUF Q6_K | 6.56 | 9.13 GB |
| GGUF Q5_K_M | 5.69 | 7.92 GB |
| GGUF Q4_K_M | 4.89 | 6.81 GB |
| GGUF Q3_K_M | 3.91 | 5.44 GB |
| GGUF Q2_K | 3.35 | 4.66 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 1× · 226 tok/s | 1× · 133 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 155 tok/s | 1× · 86.4 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 116 tok/s | 1× · 62.5 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 167 tok/s | 1× · 94.2 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 114 tok/s | 1× · 48.1 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 33.8 tok/s | 1× · 17.2 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 60.1 tok/s | 1× · 31.2 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 86.1 tok/s | 1× · 36.2 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 54.5 tok/s | 2× · 35.2 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 37.5 tok/s | 1× · 8.7 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 11.0 tok/s | 1× · 5.5 tok/s |