Instructions to use microsoft/Phi-mini-MoE-instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use microsoft/Phi-mini-MoE-instruct with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="microsoft/Phi-mini-MoE-instruct", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-mini-MoE-instruct", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("microsoft/Phi-mini-MoE-instruct", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use microsoft/Phi-mini-MoE-instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "microsoft/Phi-mini-MoE-instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/Phi-mini-MoE-instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/microsoft/Phi-mini-MoE-instruct

SGLang

How to use microsoft/Phi-mini-MoE-instruct with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "microsoft/Phi-mini-MoE-instruct" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/Phi-mini-MoE-instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "microsoft/Phi-mini-MoE-instruct" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/Phi-mini-MoE-instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use microsoft/Phi-mini-MoE-instruct with Docker Model Runner:
```
docker model run hf.co/microsoft/Phi-mini-MoE-instruct
```

cliang1453 commited on Jun 23, 2025

Commit

4bf6e2c

verified ·

1 Parent(s): bd3840a

Update README.md

Browse files

Files changed (1) hide show

README.md +10 -12

README.md CHANGED Viewed

@@ -2,6 +2,8 @@
 license: mit
 language:
 - en
 base_model:
 - microsoft/GRIN-MoE
 - microsoft/Phi-3.5-MoE-instruct
@@ -9,9 +11,11 @@ pipeline_tag: text-generation
 ---
 ## Model Summary
-The Phi-mini-MoE is a 7.6B total parameters with 2.4B activated parameters, lightweight, state-of-the-art open Mixture of Expert (MoE) model compressed and distilled from Phi-3.5-MoE using [SlimMoE](http:\\link.to.slimmoe). The training process utilizes Phi-3 datasets that includes both synthetic data and the filtered publicly available websites data with a focus on high-quality and reasoning dense properties. The model has underwent a post-training process that incorporates both supervised fine-tuning and direct preference optimization for the instruction following and safety measures. The model belongs to the SlimMoE series, where a smaller version model Phi-tiny-MoE with 3.8B total parameters and 1.1B activated parameters is available.
-📖 [SlimMoE Paper](http:\\link.to.slimmoe) <br>
 📖 [Phi-3 Technical Report](https://arxiv.org/abs/2404.14219) <br>
 📖 [GRIN-MoE](https://arxiv.org/abs/2409.12136) <br>
@@ -19,13 +23,7 @@ The Phi-mini-MoE is a 7.6B total parameters with 2.4B activated parameters, ligh
 ### Primary Use Cases
-The model is intended for commercial and research use in English. The model provides uses for general purpose AI systems and applications which require:
-1) Memory/compute constrained environments
-2) Latency bound scenarios
-3) Strong reasoning (especially math and logic)
-Our model is designed to accelerate research on language and multimodal models, for use as a building block for generative AI powered features.
 ### Use Case Considerations
@@ -90,12 +88,12 @@ print(output[0]['generated_text'])
 ## Benchmarks
-To understand the capabilities, we compare Phi-mini-MoE with a set of models over a variety of benchmarks using lm eval.
 | Model                 | # Total param | # Act. param | MMLU  | MMLU pro | BBH   | Arc-C (chat) | Human-eval | GSM8K | MT-bench |
 |----------------------|---------------|--------------|-------|----------|-------|---------------|-------------|--------|----------|
 | **MoE Models** |||||||||||
-| Phi 3.5-MoE          | 42B           | 6.6B         | 78.36 | 59.38    | 63.93 | 91.38         | 81.70       | 87.87  | 8.34     |
 | Qwen 1.5 MoE         | 14B           | 2.7B         | 60.73 | 26.49    | 42.65 | 67.24         | 46.30       | 53.07  | 6.55     |
 | DeepSeek V2 Lite     | 16B           | 2.4B         | 56.69 | 17.89    | 36.30 | 61.09         | 54.40       | 63.23  | 6.82     |
 | OL-MoE               | 7B            | 1.3B         | 54.27 | 20.87    | 38.00 | 55.63         | 37.80       | 71.49  | 6.60     |
@@ -110,7 +108,7 @@ To understand the capabilities, we compare Phi-mini-MoE with a set of models ove
 | Qwen 2.5 3B          | 3B            | 3B           | 65.06 | 41.00    | 46.61 | 80.20         | 73.80       | 76.57  | 7.60     |
 | Gemma 3 1B           | 1B            | 1B           | 40.80 | 14.70    | 34.80 | 37.46         | 41.50       | 41.77  | 6.67     |
 | LLaMA 3.2 1B         | 1B            | 1B           | 46.30 | 18.67    | 35.18 | 49.91         | 35.40       | 44.96  | 5.23     |
-| **Our (SlimMoE) Models**       |||||||||||
 | Phi-mini-MoE         | 7.6B          | 2.4B         | 70.68 | 49.68    | 55.27 | 84.91         | 73.80       | 84.89  | 7.59     |
 | Phi-tiny-MoE         | 3.8B          | 1.1B         | 60.83 | 36.34    | 45.58 | 76.37         | 58.50       | 78.47  | 7.05     |

 license: mit
 language:
 - en
+context_length:
+- 4k
 base_model:
 - microsoft/GRIN-MoE
 - microsoft/Phi-3.5-MoE-instruct
 ---
 ## Model Summary
+Phi-mini-MoE is a lightweight Mixture of Experts (MoE) model with 7.6B total parameters and 2.4B activated parameters. It is compressed and distilled from the base model shared by [Phi-3.5-MoE](https://huggingface.co/microsoft/Phi-3.5-MoE-instruct) and [GRIN-MoE](https://huggingface.co/microsoft/GRIN-MoE) using the [SlimMoE](http://link.to.slimmoe) approach, then post-trained via supervised fine-tuning and direct preference optimization for instruction following and safety. The model is trained on Phi-3 synthetic data and filtered public documents, with a focus on high-quality, reasoning-dense content. It is part of the SlimMoE series, which includes a smaller variant, [Phi-tiny-MoE](https://huggingface.co/microsoft/Phi-tiny-MoE-instruct), with 3.8B total and 1.1B activated parameters.
+References: <br>
+📖 [SlimMoE](http:\\link.to.slimmoe) <br>
 📖 [Phi-3 Technical Report](https://arxiv.org/abs/2404.14219) <br>
 📖 [GRIN-MoE](https://arxiv.org/abs/2409.12136) <br>
 ### Primary Use Cases
+The model is intended for commercial and research use in English. The model provides uses for general purpose AI systems and applications which require memory/compute constrained environments and latency bound scenarios.
 ### Use Case Considerations
 ## Benchmarks
+To understand the capabilities, we compare Phi-mini-MoE with a set of models over a variety of benchmarks using lm eval. Detailed evaluation settings can be found in the SlimMoE paper.
 | Model                 | # Total param | # Act. param | MMLU  | MMLU pro | BBH   | Arc-C (chat) | Human-eval | GSM8K | MT-bench |
 |----------------------|---------------|--------------|-------|----------|-------|---------------|-------------|--------|----------|
 | **MoE Models** |||||||||||
+| Phi-3.5-MoE          | 42B           | 6.6B         | 78.36 | 59.38    | 63.93 | 91.38         | 81.70       | 87.87  | 8.34     |
 | Qwen 1.5 MoE         | 14B           | 2.7B         | 60.73 | 26.49    | 42.65 | 67.24         | 46.30       | 53.07  | 6.55     |
 | DeepSeek V2 Lite     | 16B           | 2.4B         | 56.69 | 17.89    | 36.30 | 61.09         | 54.40       | 63.23  | 6.82     |
 | OL-MoE               | 7B            | 1.3B         | 54.27 | 20.87    | 38.00 | 55.63         | 37.80       | 71.49  | 6.60     |
 | Qwen 2.5 3B          | 3B            | 3B           | 65.06 | 41.00    | 46.61 | 80.20         | 73.80       | 76.57  | 7.60     |
 | Gemma 3 1B           | 1B            | 1B           | 40.80 | 14.70    | 34.80 | 37.46         | 41.50       | 41.77  | 6.67     |
 | LLaMA 3.2 1B         | 1B            | 1B           | 46.30 | 18.67    | 35.18 | 49.91         | 35.40       | 44.96  | 5.23     |
+| **SlimMoE Models**       |||||||||||
 | Phi-mini-MoE         | 7.6B          | 2.4B         | 70.68 | 49.68    | 55.27 | 84.91         | 73.80       | 84.89  | 7.59     |
 | Phi-tiny-MoE         | 3.8B          | 1.1B         | 60.83 | 36.34    | 45.58 | 76.37         | 58.50       | 78.47  | 7.05     |