Gemma 4 · released 2026-03 · 5.12B parameters
Per-Layer Embeddings: 2.35B of params are PLE tables (262144 x 35 x 256), 'effective' ~2B. Last 20 layers reuse KV (no own cache) and use a double-wide MLP (12288). Includes vision + audio encoders.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.02 GB | 0.01 GB |
| 8,192 tokens | 0.03 GB | 0.01 GB |
| 32,768 tokens | 0.10 GB | 0.05 GB |
| 131,072 tokens | 0.38 GB | 0.19 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 9.54 GB |
| FP8 (W8A8) | 8 | 5.15 GB |
| INT8 (W8A8 / W8A16) | 8 | 5.15 GB |
| NVFP4 | 4.5 | 3.22 GB |
| INT4 AWQ / GPTQ | 4.25 | 3.09 GB |
| GGUF Q8_0 | 8.5 | 5.07 GB |
| GGUF Q6_K | 6.56 | 3.91 GB |
| GGUF Q5_K_M | 5.69 | 3.39 GB |
| GGUF Q4_K_M | 4.89 | 2.92 GB |
| GGUF Q3_K_M | 3.91 | 2.33 GB |
| GGUF Q2_K | 3.35 | 2.00 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× · 403 tok/s | 1× · 386 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 301 tok/s | 1× · 285 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 237 tok/s | 1× · 223 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 320 tok/s | 1× · 304 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 239 tok/s | 1× · 216 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 77.6 tok/s | 1× · 71.9 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 133 tok/s | 1× · 124 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 165 tok/s | 1× · 137 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 114 tok/s | 1× · 102 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 82.2 tok/s | 1× · 20.1 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 26.2 tok/s | 1× · 24.2 tok/s |