All models › google
gemma-3n-E2B-it
google · license to accept license · on Arena since 12 Jun 2025
Arena score—Not voted on Arena yet
Size5.4 Bparameters
4-bit download3.3 GBweights only, Q4_K_M
Max context32Ktokens, from the model config
Where to use it for free
- The lab's own chat: aistudio.google.com (free with daily limits)
- Download and run it at home: official weights · GGUF files for Ollama and LM Studio (you accept the license on Hugging Face first)
No Arena votes yet
Arena has not voted on this model. It is listed because it is a popular small model on its lab's official Hugging Face page (123K downloads a month).
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | Fits in GPU | 68 tok/s |
| RTX 4060 Ti 16 GB | 16 GB | Fits in GPU | 54 tok/s |
| RTX 3090 | 24 GB | Fits in GPU | 177 tok/s |
| RTX 4090 | 24 GB | Fits in GPU | 191 tok/s |
| RTX 5090 | 32 GB | Fits in GPU | 339 tok/s |
| RX 7900 XTX | 24 GB | Fits in GPU | 182 tok/s |
| 2 × RTX 3090 | 48 GB | Fits in GPU | 172 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 162 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 103 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 48 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 524 tok/s |
Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 70% ÷ data read per token; real numbers vary by about ±30%.
GPU memory needed
| Level | 8K context | 32K context | Quality vs full |
|---|---|---|---|
| Q3_K_M | 3.4 GB | 3.7 GB | +10.5% noticeable loss |
| Q4_K_M | 4.0 GB | 4.3 GB | +2.8% small loss |
| Q5_K_M | 4.6 GB | 4.9 GB | +0.9% very small loss |
| Q6_K | 5.2 GB | 5.5 GB | +0.3% near-lossless |
| Q8_0 | 6.5 GB | 6.8 GB | ≈ 0% same as full |
| F16 | 12 GB | 12 GB | ≈ 0% same as full |
Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.