Arm/phi-3-5-mini-int8-xnnpack-executorch-raspberrypi5

Phi-3.5-mini-instruct optimized for Arm-based Edge Linux

Phi-3.5-mini-instruct text generation model, optimized with W8A8 AWQ and exported as an ExecuTorch program for the XNNPACK CPU backend on Arm-based platforms.

This model was optimized using Arm AI Model Optimizer.

Summary

This repository contains an optimized version of microsoft/Phi-3.5-mini-instruct for text generation. The model is provided as an ExecuTorch .pte file using the XNNPACK delegate, targeting Edge Linux systems. Precision: INT8 symmetric per-channel weights, INT8 dynamic activations.

Key results

Area	Result
Model format	ExecuTorch (.pte)
Target device class	Edge Linux
Reference device	Raspberry Pi 5 (16GB) (Cortex-A76, 4 cores)
Primary performance result	1.31x lower time to first token (33,931.92 ms to 25,879.87 ms, 512-token prompt)
Accuracy result	HellaSwag accuracy 66.00% (-1.00 pp vs. baseline)
Size / memory result	1.49x smaller model (7,645.04 MB to 5,135.44 MB), 5,780.06 MB peak memory

Original model

Field	Value
Original model	microsoft/Phi-3.5-mini-instruct
Original source	Hugging Face
Original developer	Microsoft
Original model card	microsoft/Phi-3.5-mini-instruct
Original license	MIT

Model files

File	Description
model.pte	Optimized INT8 ExecuTorch program for deployment
metadata.yaml	Optimized model metadata
benchmarks/	Baseline and optimized benchmark records

Performance

Performance was measured on the reference configuration below. Results do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	Raspberry Pi 5 (16GB)
CPU / accelerator	Cortex-A76, 4 cores, 2.4 GHz
OS	Raspberry Pi OS Bookworm
Runtime	ExecuTorch 1.1.0
Backend / delegate	XNNPACK
Batch size	1
Prompt length	512
Generation length	1
Precision	INT8 symmetric per-channel weights, INT8 dynamic activations
Runs	200 measured (baseline: 30 measured)

Performance results

Metric	Original / baseline	Optimized	Improvement
Time to first token (512-token prompt)	33,931.92 ms	25,879.87 ms	1.31x
Peak memory	7,668.70 MB	5,780.06 MB	1.33x less
Model size	7,645.04 MB	5,135.44 MB	1.49x smaller

On Raspberry Pi 5, latency is the time to first token for a 512-token prompt (prefill), not per-token decode latency.

Accuracy

Evaluation setup

Field	Value
Benchmark	HellaSwag
Evaluation runtime	ExecuTorch

Accuracy results

Metric	Original / baseline	Optimized	Change
HellaSwag accuracy	67.00%	66.00%	-1.00 pp

Users should re-evaluate the model on their own data before production use.

Optimization approach

Optimization area	Applied?	Notes
Model conversion	Yes	PyTorch model exported to ExecuTorch (.pte)
Quantization	Yes	INT8 symmetric per-channel weights, INT8 dynamic activations
Runtime/backend selection	Yes	XNNPACK
Accuracy validation	Yes	HellaSwag
Performance validation	Yes	Measured on Raspberry Pi 5 (16GB)

Intended use

This model is intended for developers evaluating text generation workloads on Arm-based Edge Linux platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Performance depends on the target device, runtime version, backend support, memory configuration, and system load.
• Quality was measured on HellaSwag only. Task-level behavior on other benchmarks was not evaluated.
• This repository is not a replacement for the original model documentation.

About this version

Original Model: microsoft/Phi-3.5-mini-instruct by Microsoft - Repository

Optimization/conversion: W8A8 AWQ and ExecuTorch .pte export with XNNPACK backend support for execution on Arm-based platforms.

Converted/optimized by: Nota AI, using Arm AI Model Optimizer

License: The Original Model and the Optimized Model are subject to MIT.

This repository contains a converted or optimized version of the Original Model (the "Optimized Model"). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.

No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model's behavior or intended use.

Original Model and Documentation

For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.

Licenses and Third-Party Terms

Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.

You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.

Purpose of this Release

The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.

Arm's publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.

To the fullest extent permitted by applicable law (i) the Optimized Model is provided "as is." Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.

You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.

Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.