| Model profile | ||
|---|---|---|
| Model | Size | Approximate size |
| Gemma 3 12BModel source: provider131,072 tokens max native context | 12B | 7.2 GB4-bit quantization |
| Compatible hardware Estimated speeds, not benchmark results: calculated from memory bandwidth and model size. Real results can differ significantly because there is no precise formula for deriving LLM generation speed from hardware specifications alone. | ||||
|---|---|---|---|---|
| Hardware | Memory | Bandwidth | vLLM | Estimated generation |
| DDR4 16GB Computer | 16 GBRAM | 51.2 GB/s | no | 3 tok/s4-bit quantization |
| DDR5 16GB Computer | 16 GBRAM | 89.6 GB/s | no | 6 tok/s4-bit quantization |
| DDR5 32GB Computer | 32 GBRAM | 89.6 GB/s | no | 6 tok/s4-bit quantization |
| Mac mini M4 | 24 GBUnified RAM | 120 GB/s | no | 9 tok/s4-bit quantization |
| RTX 2000 Ada | 16 GBGPU VRAM | 224 GB/s | yes | 16 tok/s4-bit quantization |
| Radeon 8060S 96GB | 96 GBUnified RAM | 256 GB/s | yes | 19 tok/s4-bit quantization |
| M4 Pro | 24 GBUnified RAM | 273 GB/s | no | 20 tok/s4-bit quantization |
| DGX Spark | 128 GBUnified RAM | 273 GB/s | yes | 20 tok/s4-bit quantization |
| RTX 4060 Ti 16GB | 16 GBGPU VRAM | 288 GB/s | yes | 21 tok/s4-bit quantization |
| M5 Pro | 64 GBUnified RAM | 307 GB/s | no | 23 tok/s4-bit quantization |
| RTX 3060 12GB | 12 GBGPU VRAM | 360 GB/s | yes | 26 tok/s4-bit quantization |
| RTX 5060 Ti 16GB | 16 GBGPU VRAM | 448 GB/s | yes | 33 tok/s4-bit quantization |
| M5 Max | 128 GBUnified RAM | 614 GB/s | no | 45 tok/s4-bit quantization |
| M3 Ultra | 96 GBUnified RAM | 819 GB/s | no | 59 tok/s4-bit quantization |
| RTX 5080 | 16 GBGPU VRAM | 960 GB/s | yes | 69 tok/s4-bit quantization |
| RTX 3090 Ti | 24 GBGPU VRAM | 1008 GB/s | yes | 72 tok/s4-bit quantization |
| M5 Ultra | 256 GBUnified RAM | 1200 GB/s | no | 85 tok/s4-bit quantization |
| RTX 5090 | 32 GBGPU VRAM | 1792 GB/s | yes | 124 tok/s4-bit quantization |
| RTX PRO 6000 | 96 GBGPU VRAM | 1792 GB/s | yes | 124 tok/s4-bit quantization |
| DGX H200 | 1128 GBGPU HBM3e | 4800 GB/s | yes | 287 tok/s4-bit quantization |
| DGX Station | 748 GBCoherent Memory | 7100 GB/s | yes | 385 tok/s4-bit quantization |
| ET900N G3 | 748 GBCoherent Memory | 7100 GB/s | yes | 385 tok/s4-bit quantization |
| DGX B200 | 1440 GBGPU HBM3e | 8000 GB/s | yes | 419 tok/s4-bit quantization |
| GB200 NVL72 | 13400 GBGPU HBM3e | 8000 GB/s | yes | 419 tok/s4-bit quantization |
| GB300 NVL72 | 20000 GBGPU HBM3e | 8000 GB/s | yes | 419 tok/s4-bit quantization |
Measures accepted-length speedup of DeepSeek's DFlash drafter on Gemma 12B using speculative decoding on a single GPU via DeepSpec resources.
Compares Gemma 4B and Gemma 12B outputs on an e-commerce app, connects Claude Desktop via MCP server, and uses DeepEval, Ragas, and Hugging Face Evaluate with LLM-as-a-judge for assessment.