Gemma 2 · released 2024-07 · 2.61B parameters
Alternating local (4096) / global layers.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.41 GB | 0.20 GB |
| 8,192 tokens | 0.61 GB | 0.30 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 4.87 GB |
| FP8 (W8A8) | 8 | 2.98 GB |
| INT8 (W8A8 / W8A16) | 8 | 2.98 GB |
| NVFP4 | 4.5 | 2.16 GB |
| INT4 AWQ / GPTQ | 4.25 | 2.10 GB |
| GGUF Q8_0 | 8.5 | 2.59 GB |
| GGUF Q6_K | 6.56 | 2.00 GB |
| GGUF Q5_K_M | 5.69 | 1.73 GB |
| GGUF Q4_K_M | 4.89 | 1.49 GB |
| GGUF Q3_K_M | 3.91 | 1.19 GB |
| GGUF Q2_K | 3.35 | 1.02 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× · 522 tok/s | 1× · 237 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 408 tok/s | 1× · 159 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 330 tok/s | 1× · 117 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 430 tok/s | 1× · 172 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 293 tok/s | 1× · 80.3 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 117 tok/s | 1× · 33.5 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 195 tok/s | 1× · 60.0 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 225 tok/s | 1× · 60.7 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 163 tok/s | 1× · 36.3 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 122 tok/s | 1× · 24.3 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 40.6 tok/s | 1× · 10.8 tok/s |