Phi-3 models
6 Phi-3 models from Microsoft on Hugging Face, from 3.8B to 14.0B parameters, published by Microsoft in BF16, with contexts from 4K to 128K tokens. The smallest official model, Phi-3-mini-128k-instruct, needs about 8.5 GB of VRAM at its published precision; the cheapest live fit is RTX 3060 at $0.110/hr.
Part of the Phi series · Previous generation: Phi-2 · Next generation: Phi-3.5
Pick a size
One row per official Phi-3 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| Phi-3-mini-4k-instruct | 3.8B | 8.5 GB | 4.3 GB | 2.1 GB | RTX 3060 | $0.110/hr | not published for this architecture |
| Phi-3-vision-128k-instruct | 4.1B | 9.3 GB | 4.6 GB | 2.3 GB | RTX 3060 | $0.110/hr | not published for this architecture |
| Phi-3-small-8k-instruct | 7.4B | 16.5 GB | 8.3 GB | 4.1 GB | RTX A5000 | $0.176/hr | 4.0 GB |
| Phi-3-medium-128k-instruct | 14.0B | 31.2 GB | 15.6 GB | 7.8 GB | RTX A6000 | $0.363/hr | not published for this architecture |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (6)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| Phi-3-mini-128k-instruct | 3.8B | BF16 | 8.5 GB | RTX 3060 | $0.110/hr |
| Phi-3-mini-4k-instruct | 3.8B | BF16 | 8.5 GB | RTX 3060 | $0.110/hr |
| Phi-3-vision-128k-instruct | 4.1B | BF16 | 9.3 GB | RTX 3060 | $0.110/hr |
| Phi-3-small-8k-instruct | 7.4B | BF16 | 16.5 GB | RTX A5000 | $0.176/hr |
| Phi-3-medium-128k-instruct | 14.0B | BF16 | 31.2 GB | RTX A6000 | $0.363/hr |
| Phi-3-medium-4k-instruct | 14.0B | BF16 | 31.2 GB | RTX A6000 | $0.363/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.