---
license: mit
language:
- en
base_model:
- unsloth/phi-4
- microsoft/phi-4
pipeline_tag: text-generation
---

# Phi-4 converted for ExLlamaV2 

[ExLlamaV2 is an inference library for running local LLMs on modern consumer GPUs.](https://github.com/turboderp-org/exllamav2)


|  | Quant type | File Size | Vram*|
| -------- | ---------- | --------- | -------- |
| [phi-4 hb8 3bpw](https://huggingface.co/cmh/phi-4_exl2/tree/hb8_3bpw) | 3 bits per weight | 6.66 GB | **10,3 GB** |
| [phi-4 hb8 4bpw](https://huggingface.co/cmh/phi-4_exl2/tree/hb8_4bpw) | 4 bits per weight | 8.36 GB | **11,9 GB** |
| [phi-4 hb8 5bpw](https://huggingface.co/cmh/phi-4_exl2/tree/hb8_5bpw) | 5 bits per weight | 10.1 GB | **13,5 GB** |
| [phi-4 hb8 6bpw](https://huggingface.co/cmh/phi-4_exl2/tree/hb8_6bpw) | 6 bits per weight | 11.8 GB | **15,1 GB** |
| [phi-4 hb8 7bpw](https://huggingface.co/cmh/phi-4_exl2/tree/hb8_7bpw) | 7 bits per weight | 13.5 GB | **16,7 GB** |
| [phi-4 hb8 8bpw](https://huggingface.co/cmh/phi-4_exl2/tree/hb8_8bpw) | 8 bits per weight | 15.2 GB | **18,2 GB** |

<sub>*approximate value at **16k context, FP16 cache**.<sup>

---------------------------------------------

# Phi-4 Model Card 

[Phi-4 Technical Report](https://arxiv.org/pdf/2412.08905)

## Model Summary 

|                         |                                                                               |     
|-------------------------|-------------------------------------------------------------------------------|
| **Developers**          | Microsoft Research                                                            |
| **Description**         | `phi-4` is a state-of-the-art open model built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets. The goal of this approach was to ensure that small capable models were trained with data focused on high quality and advanced reasoning.<br><br>`phi-4` underwent a rigorous enhancement and alignment process, incorporating both supervised fine-tuning and direct preference optimization to ensure precise instruction adherence and robust safety measures                |
| **Architecture**        | 14B parameters, dense decoder-only Transformer model                          |
| **Context length**      | 16384 tokens                                                                    |

## Usage

### Input Formats

Given the nature of the training data, `phi-4` is best suited for prompts using the chat format as follows: 

```bash
<|im_start|>system<|im_sep|>
You are a medieval knight and must provide explanations to modern people.<|im_end|>
<|im_start|>user<|im_sep|>
How should I explain the Internet?<|im_end|>
<|im_start|>assistant<|im_sep|>
```

### With ExUI:

Add Phi-4 prompt format:

Edit/replace exui/backend/prompts.py with https://huggingface.co/cmh/phi-4_exl2/raw/main/backend/prompts.py