Gemma 3 · released 2025-03 · 4.30B parameters
5 local (1024) : 1 global. params include ~0.4B SigLIP vision encoder.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.19 GB | 0.10 GB |
| 8,192 tokens | 0.27 GB | 0.13 GB |
| 32,768 tokens | 0.74 GB | 0.37 GB |
| 131,072 tokens | 2.61 GB | 1.31 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 8.01 GB |
| FP8 (W8A8) | 8 | 4.63 GB |
| INT8 (W8A8 / W8A16) | 8 | 4.63 GB |
| NVFP4 | 4.5 | 3.15 GB |
| INT4 AWQ / GPTQ | 4.25 | 3.05 GB |
| GGUF Q8_0 | 8.5 | 4.25 GB |
| GGUF Q6_K | 6.56 | 3.28 GB |
| GGUF Q5_K_M | 5.69 | 2.85 GB |
| GGUF Q4_K_M | 4.89 | 2.45 GB |
| GGUF Q3_K_M | 3.91 | 1.96 GB |
| GGUF Q2_K | 3.35 | 1.68 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× · 422 tok/s | 1× · 295 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 318 tok/s | 1× · 206 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 251 tok/s | 1× · 156 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 338 tok/s | 1× · 222 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 235 tok/s | 1× · 121 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 83.5 tok/s | 1× · 46.7 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 142 tok/s | 1× · 82.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 178 tok/s | 1× · 91.4 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 124 tok/s | 1× · 57.2 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 90.8 tok/s | 1× · 23.6 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 28.3 tok/s | 1× · 15.3 tok/s |