Phi-4 models
7 Phi-4 models from Microsoft on Hugging Face, from 3.8B to 14.7B parameters, published by Microsoft in BF16, with contexts from 16K to 128K tokens. The smallest official model, Phi-4-mini-instruct, needs about 8.6 GB of VRAM at its published precision; the cheapest live fit is RTX 4070 at $0.154/hr.
Part of the Phi series · Previous generation: Phi-3.5
Pick a size
One row per official Phi-4 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-4-mini-instruct | 3.8B | 8.6 GB | 4.3 GB | 2.1 GB | RTX 4070 | $0.154/hr | not published for this architecture |
| Phi-4-multimodal-instruct | 5.6B | 12.5 GB | 6.2 GB | 3.1 GB | RTX 3090 | $0.162/hr | not published for this architecture |
| phi-4 | 14.7B | 32.8 GB | 16.4 GB | 8.2 GB | A100 | $0.462/hr | 6.3 GB |
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-4-mini-instruct | 3.8B | BF16 | 8.6 GB | RTX 4070 | $0.154/hr |
| Phi-4-mini-reasoning | 3.8B | BF16 | 8.6 GB | RTX 4070 | $0.154/hr |
| Phi-4-multimodal-instruct | 5.6B | BF16 | 12.5 GB | RTX 3090 | $0.162/hr |
| phi-4 | 14.7B | BF16 | 32.8 GB | A100 | $0.462/hr |
| Phi-4-reasoning | 14.7B | BF16 | 32.8 GB | A100 | $0.462/hr |
| Phi-4-reasoning-plus | 14.7B | BF16 | 32.8 GB | A100 | $0.462/hr |
Quantized versions (1)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| phi-4-FP8-dynamic | 14.7B | F8_E4M3 | 16.4 GB | RTX 4000 Ada | $0.347/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.
Run and fine-tune Phi-4
- How to run Phi-4: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune Phi-4: LoRA and QLoRA memory per model, and the license limits on derivatives.
