Phi-3.5-mini-instruct-onnx

Phi-3.5-Mini-Instruct ONNX models

This repository hosts the optimized versions of Phi-3.5-mini-instruct to accelerate inference with ONNX Runtime.

Optimized Phi-3.5 Mini models are published here in ONNX format to run with ONNX Runtime on CPU and GPU across devices, including server platforms, Windows, Linux and Mac desktops, and mobile CPUs, with the precision best suited to each of these targets.

To easily get started with Phi-3.5, you can use our newly introduced ONNX Runtime Generate() API. See here for instructions on how to run it.

ONNX Models

Here are some of the optimized configurations we have added:

1. ONNX model for INT4 CPU: ONNX model for CPUs using int4 quantization via AWQ.
2. ONNX model for INT4 GPU: ONNX model for GPUs using int4 quantization via AWQ.

Model Summary

Phi-3.5 mini is a lightweight, state-of-the-art open model built upon datasets used for Phi-3 - synthetic data and filtered publicly available websites - with a focus on very high-quality, reasoning dense data. The model belongs to the Phi-3 model family and supports 128K token context length. The model underwent a rigorous enhancement process, incorporating both supervised fine-tuning, proximal policy optimization, and direct preference optimization to ensure precise instruction adherence and robust safety measures.

Intended Uses

The Phi 3.5 mini model is intended for commercial and research use in multiple languages. 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 code, math and logic)

Use Case Considerations

Phi 3.5 models are not specifically designed or evaluated for all downstream purposes. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fariness before using within a specific downstream use case, particularly for high risk scenarios. Developers should be aware of and adhere to applicable laws or regulations (including privacy, trade compliance laws, etc.) that are relevant to their use case.

Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.

Release Notes

This is an update over the instruction-tuned Phi-3 Mini ONNX model release. We believe most use cases will benefit from this release, but we encourage users to test their particular AI applications. We appreciate the enthusiastic adoption of the Phi-3 model family and continue to welcome all feedback from the community.

What’s New (2026-02)

This update introduces an improved INT4 GPU ONNX model that incorporates quantization-aware fine-tuning (QAT) on top of the existing quantization pipeline.

Benchmark Accuracy Improvements (INT4 GPU)

The table above provides a high-level summary of observed accuracy deltas across benchmark categories compared to the old INT4 GPU model. The QAT-tuned INT4 GPU model improves performance on the majority of downstream reasoning and QA benchmarks, with a small regression on broad-coverage evaluation.

Generation Stability (EOS Behavior)

The updated model reduces EOS non-emission by approximately 5× compared to the previous INT4 GPU ONNX release, as observed across a large set of randomly generated prompts, resulting in more reliable sequence termination and generation behavior closer to the Torch baseline.

Hardware Supported

The ONNX models are tested on:

• GPU SKU: RTX 4090 (DirectML)
• GPU SKU: 1 A100 80GB GPU, SKU: Standard_ND96amsr_A100_v4 (CUDA)
• CPU SKU: Standard D16s v6 (16 vcpus, 64 GiB memory)
• AMD CPU: Internal_D64as_v5

Minimum Configuration Required:

• Windows: DirectX 12-capable GPU and a minimum of 4GB of combined RAM
• CUDA: NVIDIA GPU with Compute Capability >= 7.0

Model Description

• Developed by: Microsoft
• Model type: ONNX
• Language(s) (NLP): Python, C, C++
• License: MIT
• Model Description: This is a conversion of the Phi-3.5 Mini-Instruct model for ONNX Runtime inference.

How to Get Started with the Model

To make running of the Phi-3.5 models across a range of devices and platforms across various execution provider backends possible, we introduce a new API to wrap several aspects of generative AI inferencing. This API make it easy to drag and drop LLMs straight into your app. For running the early version of these models with ONNX Runtime, follow the steps here.

For example:

python model-qa.py -m /*{YourModelPath}*/Phi-3.5-mini-instruct-onnx/cpu_and_mobile/cpu-int4-awq-block-128-acc-level-4 -k 40 -p 0.95 -t 0.8 -r 1.0

*Input:*  <|user|>Tell me a joke<|end|><|assistant|>

*Output:*  Why don't scientists trust atoms?
           Because they make up everything!

This joke plays on the double meaning of "make up." In science, atoms are the fundamental building blocks of matter, literally making up everything. However, in a colloquial sense, "to make up" can mean to fabricate or lie, hence the humor.

Performance Metrics

Phi-3.5 Mini-Instruct performs better in ONNX Runtime than PyTorch for all batch size, prompt length combinations.

The table below shows the average throughput of the first 256 tokens generated (tps) for FP16 and INT4 precisions on CUDA as measured on 1 A100 80GB GPU, SKU: Standard_ND96amsr_A100_v4. ONNX Runtime models for GPU are 21X faster than PyTorch Compile and up to 8X faster than llama.cpp on A100 GPU.

The table below shows the average throughput of the first 256 tokens generated (tps) for INT4 precision on CPU as measured on a Standard D16s v6 (16 vcpus, 64 GiB memory)

Package Versions

Appendix

Activation Aware Quantization (AWQ) works by identifying the top 1% most salient weights that are most important for maintaining accuracy and quantizing the remaining 99% of weights. This leads to less accuracy loss from quantization compared to many other quantization techniques. For more on AWQ, see here.

Model Card Contact

parinitarahi

Contributors

Sunghoon Choi, Yufeng Li, Kunal Vaishnavi, Akshay Sonawane, Rui Ren, Parinita Rahi, Nenad Banfic

License

The model is licensed under the MIT license.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft’s Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party’s policies.