Arm/gemma-3-1b-instruct-onnx-genai-int4-emb-int8

Gemma-3-1B-IT optimized for Arm-based mobile CPUs with SME2

A decoder-only instruction-tuned language model for text generation, provided in ONNX for the ONNX Runtime GenAI runtime and quantized to INT4 groupwise asymmetric weights with INT8 per-token dynamic activations and an INT8 block-quantized embedding table tied to the INT8 LM head, targeting Arm-based mobile CPUs with SME2.

Summary

This repository contains an Arm-optimized version of google/gemma-3-1b-it for text generation. The model is provided in ONNX for the ONNX Runtime GenAI runtime, targeting Mobile CPU systems. The transformer weights are quantized with GPTQ to INT4, non-symmetric per-group with group size 32, activations are INT8 per-token dynamic, and the embedding table is INT8 block-quantized with block size 32 and tied to the INT8 LM head (scheme W4A8_dyn_emb_int8_tied).

This is the instruction-tuned checkpoint. It ships a chat template and expects prompts wrapped in the Gemma 3 instruct turn markers.

This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on MMLU and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	ONNX
Target device class	Mobile CPU
Reference device	vivo X300 (C1-Ultra, C1-Premium, C1-Pro, Android 16 / OriginOS 6)
Primary performance result	35.05 tokens/sec decode throughput (4.10x the FP32 baseline); 633.21 ms time to first token, which is 30.5 % faster than the FP32 baseline
Accuracy result	MMLU 5-shot 37.84 %
Size / memory result	825.38 MB with INT4 weights, INT8 activations and an INT8 embedding table tied to the INT8 LM head (4.70x smaller than the FP32 baseline); peak memory 1501.98 MB

Original model

Field	Value
Original model	google/gemma-3-1b-it
Original source	Hugging Face
Original developer	Google
Original model card	google/gemma-3-1b-it
Original license	Gemma Terms of Use

Model files

File	Description
model.onnx	Arm-optimized INT4-weight model graph for deployment; weights ship in the model.onnx.data side-car
example.py	Minimal inference example
pyproject.toml	Pinned runtime dependencies for example.py, resolved with uv
uv.lock	Locked dependency resolution for pyproject.toml
config.yaml	Model I/O contract used by the example
benchmarks/	FP32 baseline and Arm-optimized benchmark records

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	vivo X300
CPU / accelerator	C1-Ultra, C1-Premium, C1-Pro, aarch64, CPU execution backend
OS	android, Android 16 / OriginOS 6
Runtime	ONNX Runtime 1.27.0
Backend / delegate	MLAS, KleidiAI
Batch size	1
Precision	INT4 groupwise non-symmetric weights (GPTQ, group size 32), INT8 per-token dynamic activations, INT8 block-quantized embedding table with block size 32 tied to the INT8 LM head (scheme W4A8_dyn_emb_int8_tied)
Threads	4 intra-op threads
Runs	5 warmup runs + 20 measured runs

Measurement conditions. Each measured run consumes a prompt of ~128 input tokens and generates ~128 output tokens, using 4 CPU threads, 5 warm-up runs and 20 measured runs. Each run starts only once Android reports thermal status 0 (NONE), after a 30 s settle. The device is set to fixed performance mode, which is the official recommendation.

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
End-to-end latency p50 (ms)	15860.39	4252.03	3.73x faster
End-to-end latency p90 (ms)	15923.09	4613.95	3.45x faster
End-to-end latency p99 (ms)	16054.16	4666.65	3.44x faster
Decode throughput (tokens/sec)	8.55	35.05	4.10x
Time to first token (ms)	910.62	633.21	1.44x faster
Model load time (ms)	9845.75	1765.63	5.58x faster
Model size (MB)	3878.56	825.38	4.70x smaller
Peak memory (MB)	5303.11	1501.98	3.53x less
Average memory (MB)	5172.10	1367.77	3.78x less

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Field	Value
Dataset	MMLU
Split	5-shot
Number of samples	8258
Metric(s)	MMLU accuracy
Evaluation runtime	ONNX Runtime 1.27.0

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
MMLU accuracy (%)	39.90	37.84	-2.06 pp

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.

For this release, Arm used:

Optimization area	Applied?	Notes
Model conversion	Yes	Converted to ONNX in the ONNX Runtime GenAI graph layout, with attention fused into a single GroupQueryAttention op and the KV cache wired internally
Quantization	Yes	GPTQ weight-only quantization to INT4, non-symmetric per-group with group size 32, calibrated on 512 WikiText-2 raw samples; INT8 per-token dynamic activations; INT8 block-quantized embedding table with block size 32, tied to the INT8 LM head (scheme W4A8_dyn_emb_int8_tied). The LM head is held at INT8 rather than INT4
Runtime/backend selection	Yes	ONNX Runtime 1.27.0 CPU execution provider, with MLAS and KleidiAI kernels
Graph/runtime compatibility updates	Yes	Performed as part of the shared PT2E export pipeline, with ONNX-specific graph translation
Accuracy validation	Yes	Compared against the original model or published baseline
Performance validation	Yes	Measured on the reference Arm platform

The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.

Using this model

Note: The Python/uv example runs on AWS Graviton (Ubuntu arm64) to confirm runtime compatibility only, and is intended as a guideline for building an equivalent run script on Mobile CPU systems. Every performance and accuracy figure on this page was measured on the vivo X300; none of them was re-measured on the proxy, and a run on the proxy is not evidence about the phone.

Install dependencies

Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:

uv python install
uv sync --frozen

Run the example

uv run example.py

The example writes predictions.json next to example.py.

Expected input

Property	Value
Input shape	[1, T], where T is the runtime sequence length, variable up to 4096 as the bundle is configured
Input type	int64, input_ids
Input range	Token ids in the range 0 to 262143, for a vocabulary size of 262144
Preprocessing	Apply the Gemma 3 instruct chat template to the user prompt, then tokenize with tokenizer.json without adding special tokens, since the template already carries the turn markers and the runtime prepends the beginning-of-sequence token. position_ids is not exposed as a graph input, and KV-cache I/O is wired up internally by the ONNX Runtime GenAI Generator.

Expected output

Property	Value
Output shape	One token id per decode step, streamed by the ONNX Runtime GenAI Generator
Output type	Token ids, integer indices into the 262144-entry vocabulary
Postprocessing	Decode with tokenizer.json, stopping on the EOS token id 1, the end-of-turn token id 106, or when max_length is reached

Intended use

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

Limitations

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• Accuracy was evaluated on MMLU 5-shot and may not generalize to all domains. It does not measure code, math, multi-turn instruction following, or long-context behavior.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.
• This is the instruction-tuned checkpoint. It expects prompts wrapped in the Gemma 3 instruct turn markers; feeding it raw completion-style text without the chat template degrades generation quality.

Additional notes

• The ONNX Runtime GenAI bundle requires the literal file names model.onnx and model.onnx.data, because genai_config.json hard-references them.
• The Gemma 3 architecture supports a longer context than this bundle exposes. The bundled genai_config.json sets the ONNX Runtime GenAI context length to 4096, trimmed for the smartphone KV-cache budget, so 4096 is the limit the runtime enforces as shipped. Raise it there if a longer context is needed.
• The quantization recipe is selective: the final LM-head projection is kept at INT8 rather than INT4, so some layers were skipped from INT4 quantization to keep accuracy within acceptable ranges. Normalization layers, residual adds, and the activation function remain in floating point inside the fused GroupQueryAttention and SkipSimplifiedLayerNormalization ops.
• The embedding table and the LM head share a single INT8 block-quantized weight tensor, with block size 32, and its scales and zero points.
• Weight quantization was calibrated with GPTQ Hessians computed on 512 WikiText-2 raw samples. Activations are quantized dynamically at runtime and need no offline calibration.
• Accuracy evaluation scope is MMLU 5-shot only.

About this version

Original Model: google/gemma-3-1b-it by Google - Repository

Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.

Converted/optimized by: Arm

License: The Original Model and the Optimized Model are subject to Gemma Terms of Use.

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.