Gemma 4 · released 2026-03 · 25.8B parameters (3.82B active per token)
MoE: every layer runs a dense MLP (2112, modelled as 1 shared expert) in parallel with 8 of 128 routed experts (704). Global layers: 2 KV heads x 512, K=V. params include ~0.57B vision encoder.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.35 GB | 0.18 GB |
| 8,192 tokens | 0.51 GB | 0.25 GB |
| 32,768 tokens | 1.45 GB | 0.72 GB |
| 131,072 tokens | 5.20 GB | 2.60 GB |
| 262,144 tokens | 10.2 GB | 5.10 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 48.1 GB |
| FP8 (W8A8) | 8 | 24.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 24.7 GB |
| NVFP4 | 4.5 | 14.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 13.8 GB |
| GGUF Q8_0 | 8.5 | 25.5 GB |
| GGUF Q6_K | 6.56 | 19.7 GB |
| GGUF Q5_K_M | 5.69 | 17.1 GB |
| GGUF Q4_K_M | 4.89 | 14.7 GB |
| GGUF Q3_K_M | 3.91 | 11.7 GB |
| GGUF Q2_K | 3.35 | 10.1 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× · 445 tok/s | 1× · 118 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 336 tok/s | 1× · 74.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 266 tok/s | 1× · 53.5 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 357 tok/s | 1× · 81.7 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 274 tok/s | 1× · 48.4 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 88.9 tok/s | 1× · 14.5 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 151 tok/s | 1× · 26.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 175 tok/s | 1× · 35.2 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 120 tok/s | 2× · 36.4 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 87.1 tok/s | 1× · 13.6 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 30.2 tok/s | 1× · 4.6 tok/s |