Gemma 3n · released 2025-06 · 5.44B parameters
2.01B of params are Per-Layer-Embedding tables (262144 x 30 x 256), offloadable; 'effective' ~2B. Last 10 layers reuse KV. Includes vision + audio encoders (~1.0B together).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.05 GB | 0.02 GB |
| 8,192 tokens | 0.08 GB | 0.04 GB |
| 32,768 tokens | 0.27 GB | 0.13 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 10.1 GB |
| FP8 (W8A8) | 8 | 5.57 GB |
| INT8 (W8A8 / W8A16) | 8 | 5.57 GB |
| NVFP4 | 4.5 | 3.57 GB |
| INT4 AWQ / GPTQ | 4.25 | 3.43 GB |
| GGUF Q8_0 | 8.5 | 5.38 GB |
| GGUF Q6_K | 6.56 | 4.15 GB |
| GGUF Q5_K_M | 5.69 | 3.60 GB |
| GGUF Q4_K_M | 4.89 | 3.10 GB |
| GGUF Q3_K_M | 3.91 | 2.48 GB |
| GGUF Q2_K | 3.35 | 2.12 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× · 396 tok/s | 1× · 354 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 291 tok/s | 1× · 254 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 226 tok/s | 1× · 195 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 310 tok/s | 1× · 272 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 223 tok/s | 1× · 177 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 72.3 tok/s | 1× · 60.4 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 125 tok/s | 1× · 105 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 162 tok/s | 1× · 130 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 110 tok/s | 1× · 84.6 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 78.8 tok/s | 1× · 19.3 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 24.2 tok/s | 1× · 20.0 tok/s |