Gemma 4 · released 2026-03 · 31.3B parameters
5 sliding (1024) : 1 global pattern; global layers use 4 KV heads x 512 head_dim and attention_k_eq_v (no v_proj, V derived from K; HF cache still stores K and V). params include ~0.57B vision encoder (text-only ~30.70B).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 1.41 GB | 0.70 GB |
| 8,192 tokens | 2.03 GB | 1.02 GB |
| 32,768 tokens | 5.78 GB | 2.89 GB |
| 131,072 tokens | 20.8 GB | 10.4 GB |
| 262,144 tokens | 40.8 GB | 20.4 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 58.3 GB |
| FP8 (W8A8) | 8 | 30.4 GB |
| INT8 (W8A8 / W8A16) | 8 | 30.4 GB |
| NVFP4 | 4.5 | 18.3 GB |
| INT4 AWQ / GPTQ | 4.25 | 17.4 GB |
| GGUF Q8_0 | 8.5 | 30.9 GB |
| GGUF Q6_K | 6.56 | 23.9 GB |
| GGUF Q5_K_M | 5.69 | 20.7 GB |
| GGUF Q4_K_M | 4.89 | 17.8 GB |
| GGUF Q3_K_M | 3.91 | 14.2 GB |
| GGUF Q2_K | 3.35 | 12.2 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× · 109 tok/s | 1× · 59.2 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 69.9 tok/s | 1× · 36.9 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 50.4 tok/s | 1× · 26.2 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 76.3 tok/s | 1× · 40.5 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 52.2 tok/s | 1× · 20.3 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 13.8 tok/s | 2× · 12.9 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 25.0 tok/s | 1× · 12.7 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 38.6 tok/s | 4× · 34.3 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 23.0 tok/s | 4× · 23.3 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 15.3 tok/s | 1× · 3.4 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 4.4 tok/s | 1× · 2.2 tok/s |