Loading…
Loading…
Models / gemma-4 / gemma-4 12B
Any to any · served with vLLM · ↓ 73K downloads a month · View on Hugging Face
The same model packaged differently: a quantized version needs less GPU memory and runs on a cheaper machine.
| Version | Runtime | Weights | GPU memory | Cheapest machine now |
|---|---|---|---|---|
| Originalgoogle/gemma-4-12B | vLLM | 24 GB | 32.4 GB | $0.44/h 2x RTX 3090 |
| itgoogle/gemma-4-12B-it | vLLM | 24 GB | 32.4 GB | $0.44/h 2x RTX 3090 |
| it qat w4a16 ctgoogle/gemma-4-12B-it-qat-w4a16-ct | vLLM | 10 GB | 17.2 GB | $0.18/h 2x RTX 3060 |
| it qat q4 0 unquantizedgoogle/gemma-4-12B-it-qat-q4_0-unquantized | vLLM | 24 GB | 32.4 GB | $0.44/h 2x RTX 3090 |
| it AWQ INT4 (cyankiwi)cyankiwi/gemma-4-12B-it-AWQ-INT4 | vLLM | 11 GB | 18.3 GB | $0.18/h 2x RTX 3060 |
| it qat AWQ INT4 (cyankiwi)cyankiwi/gemma-4-12B-it-qat-AWQ-INT4 | vLLM | 11 GB | 18.3 GB | $0.18/h 2x RTX 3060 |
| it NVFP4 (RedHatAI)RedHatAI/gemma-4-12B-it-NVFP4 | vLLM | 10 GB | 17.2 GB | $0.18/h 2x RTX 3060 |
| OBLITERATED (OBLITERATUS)OBLITERATUS/Gemma-4-12B-OBLITERATED | vLLM | 24 GB | 32.4 GB | $0.44/h 2x RTX 3090 |
| it FP8 Dynamic (RedHatAI)RedHatAI/gemma-4-12B-it-FP8-Dynamic | vLLM | 15 GB | 22.5 GB | $0.23/h RTX A5000 |
| it FP8 dynamic (coolthor)coolthor/gemma-4-12B-it-FP8-dynamic | vLLM | 13 GB | 20.3 GB | $0.18/h 2x RTX 3060 |
| AriannaV4 (SurgeFF)SurgeFF/AriannaV4 | vLLM | 24 GB | 32.4 GB | $0.44/h 2x RTX 3090 |
| it qat w4a16 (unsloth)unsloth/gemma-4-12B-it-qat-w4a16 | vLLM | 10 GB | 17.2 GB | $0.18/h 2x RTX 3060 |
$0.61/h from your balance while it runs, billed by the hour; stop it from its page. GridPup rents the machine, which downloads the model from Hugging Face (usually ready in 10 to 15 minutes), and gives you an https address, keys and a Playground.
Prices are per hour for the whole machine and its disk, right now. Speed is Vast's DLPerf score of the machine, relative to the cheapest. Best value: the most speed per dollar. Only machines with a Vast reliability above 95% that download at 400 Mbit/s or more are offered: the machine fetches the model itself.