Gemma 3 · released 2025-03 · 1.00B parameters
Text-only. 5 local (512) : 1 global.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.03 GB | 0.01 GB |
| 8,192 tokens | 0.04 GB | 0.02 GB |
| 32,768 tokens | 0.14 GB | 0.07 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 1.86 GB |
| FP8 (W8A8) | 8 | 1.21 GB |
| INT8 (W8A8 / W8A16) | 8 | 1.21 GB |
| NVFP4 | 4.5 | 0.93 GB |
| INT4 AWQ / GPTQ | 4.25 | 0.91 GB |
| GGUF Q8_0 | 8.5 | 0.99 GB |
| GGUF Q6_K | 6.56 | 0.76 GB |
| GGUF Q5_K_M | 5.69 | 0.66 GB |
| GGUF Q4_K_M | 4.89 | 0.57 GB |
| GGUF Q3_K_M | 3.91 | 0.46 GB |
| GGUF Q2_K | 3.35 | 0.39 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× · 708 tok/s | 1× · 636 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 620 tok/s | 1× · 532 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 546 tok/s | 1× · 452 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 638 tok/s | 1× · 554 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 500 tok/s | 1× · 381 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 256 tok/s | 1× · 186 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 382 tok/s | 1× · 293 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 335 tok/s | 1× · 262 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 282 tok/s | 1× · 199 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 237 tok/s | 1× · 85.0 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 101 tok/s | 1× · 68.5 tok/s |