Hardware to run Gemma 4 E4B
Google's on-device model: about 4B effective parameters per token, with image and audio input.
- Developer
- Parameters
- 8B (active 4B)
- Max context
- 128K tokens
- Licence
- Apache 2.0
Memory needed
Factual dataWeights plus context cache plus about 1.5 GB of runtime buffers, by quantisation and context length.
| Quantisation | 8K tokens | 32K tokens | 128K tokens |
|---|---|---|---|
| Q4_K_M | 6.8 GB | 7.5 GB | 10 GB |
| Q8_0 | 10 GB | 11 GB | 13 GB |
| FP16 | 18 GB | 18 GB | 21 GB |
Hardware that runs it
Estimate30 platforms can run Gemma 4 E4B at Q4_K_M with a 32K-token context. Needs about 7.5 GB.
Cheapest that runs it
Fastest that runs it
Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)
RunsRuns comfortably (30)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| RTX PRO 6000 | 96 GB | ≤ 720 tok/s | $8,565 | |
| RTX 5090 | 32 GB | ≤ 720 tok/s | $4,329 | |
| RTX PRO 5000 | 48 GB | ≤ 540 tok/s | $4,500 | |
| Mac Studio M5 Ultra | 96 GB | ≤ 482 tok/s | $5,499 | |
| RTX 4090 | 24 GB | ≤ 405 tok/s | $1,599 | |
| RX 7900 XTX | 24 GB | ≤ 386 tok/s | $1,358 | |
| RTX 5080 | 16 GB | ≤ 386 tok/s | $1,830 | |
| RTX 3090 | 24 GB | ≤ 376 tok/s | $1,340 | |
| RTX PRO 4500 | 32 GB | ≤ 360 tok/s | $2,500 | |
| RTX 5070 Ti | 16 GB | ≤ 360 tok/s | $1,190 | |
| RTX PRO 4000 | 24 GB | ≤ 270 tok/s | $1,500 | |
| RTX 5070 | 12 GB | ≤ 270 tok/s | $1,010 | |
| Radeon AI PRO R9700 | 32 GB | ≤ 259 tok/s | $1,550 | |
| RX 9070 XT | 16 GB | ≤ 259 tok/s | $759 | |
| Arc Pro B70 | 32 GB | ≤ 244 tok/s | $1,268 | |
| Mac Studio M5 Max | 36 GB | ≤ 185 tok/s | $2,499 | |
| Arc Pro B60 | 24 GB | ≤ 183 tok/s | $599 | |
| Arc B580 | 12 GB | ≤ 183 tok/s | $249 | |
| RTX 5060 Ti 16 GB | 16 GB | ≤ 180 tok/s | $789 | |
| RTX 3060 12 GB | 12 GB | ≤ 145 tok/s | $329 | |
| RX 9060 XT 16 GB | 16 GB | ≤ 129 tok/s | $349 | |
| Mac mini M5 Pro | 24 GB | ≤ 123 tok/s | $1,699 | |
| Ryzen AI Max+ PRO 495 | 192 GB | ≤ 110 tok/s | $7,399 | |
| Jetson AGX Thor | 128 GB | ≤ 110 tok/s | $5,499 | |
| DGX Spark / GB10 | 128 GB | ≤ 110 tok/s | $5,000 | |
| Ryzen AI Max+ 395 | 64 GB | ≤ 103 tok/s | $2,200 | |
| Arc Pro B50 | 16 GB | ≤ 90 tok/s | $349 | |
| Jetson AGX Orin 64GB | 64 GB | ≤ 82 tok/s | $1,999 | |
| Mac mini M6 | 16 GB | ≤ 68 tok/s | $899 | |
| Ryzen AI 9 HX 370 | 32 GB | ≤ 48 tok/s | — |
Not enough memoryNot enough memory (1)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| Jetson Orin Nano Super | 8 GB | — | $399 |
Not supportedNot supported (2)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| RK3588 (Orange Pi 5 Plus and others) | 32 GB | — | — | — |
| Raspberry Pi AI HAT+ 2 | 8 GB | — | — | $200 |
Theoretical maximum for generating text: memory bandwidth divided by the bytes read per token. Real runs usually reach about half to three quarters of it.