Arm/llama-3-2-1b-instruct-onnx-genai-int4-kquantlast-emb-int8-vivo-x300

Llama-3.2-1B-Instruct optimized for Arm-based mobile CPUs with SME2

A quantized version of meta-llama/Llama-3.2-1B-Instruct for text generation, provided in ONNX for the ONNX Runtime GenAI runtime with INT4 weights, dynamic INT8 activations and INT8 token embeddings, and optimized for Arm-based mobile CPUs with SME2.

Summary

This repository contains an Arm-optimized version of meta-llama/Llama-3.2-1B-Instruct for text generation. The model is provided in ONNX for the ONNX Runtime GenAI runtime, targeting Mobile CPU systems.

The weights are quantized with GPTQ to 4-bit asymmetric per-group values (group size 32), activations are quantized to INT8 per-token dynamic, and the token-embedding table and LM head weights are kept in INT8. The composite scheme is labelled W4A8_dyn_emb_int8. The graph is packed for the ONNX Runtime GenAI runtime, which fuses attention into a single GroupQueryAttention operator and wires up the KV cache.

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.

This is the instruction-tuned checkpoint. The bundled chat template is applied to the prompt before tokenization, so the model answers chat-style turns. Decode throughput, latency, time to first token, model size and memory all improve over the FP32 baseline.

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	30.56 tokens/sec decode, 949.98 ms time to first token
Accuracy result	MMLU 44.2%
Size / memory result	1167.52 MB model size (4.07x smaller than the FP32 baseline), peak memory 1280.24 MB

Original model

Field	Value
Original model	meta-llama/Llama-3.2-1B-Instruct
Original source	Hugging Face
Original developer	Meta
Original model card	meta-llama/Llama-3.2-1B-Instruct
Original license	Llama 3.2 Community License

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 (CPU execution backend)
OS	android, Android 16 / OriginOS 6
Runtime	ONNX Runtime 1.27.0
Backend / delegate	MLAS, KleidiAI
Batch size	1
Threads	4 intra-op threads
Precision	INT4 weights (GPTQ, asymmetric, per-group, group size 32), dynamic per-token INT8 activations, INT8 token embeddings
Runs	5 warmup runs + 20 measured runs
Prompt / generation length	162 prompt tokens, 128 generated tokens

Measurement conditions. Each measured run consumes a 162-token prompt and generates 128 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
p50 latency	19230.35 ms	5161.69 ms	3.73x faster
p90 latency	19315.52 ms	5220.15 ms	3.70x faster
p99 latency	19477.92 ms	5233.12 ms	3.72x faster
Time to first token	1258.15 ms	949.98 ms	1.32x faster
Decode throughput	7.10 tokens/sec	30.56 tokens/sec	4.30x
Model load time	10284.67 ms	1606.96 ms	6.40x faster
Model size	4746.40 MB	1167.52 MB	4.07x smaller
Peak memory	5974.20 MB	1280.24 MB	4.67x less
Average memory	5895.43 MB	1196.22 MB	4.93x 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 (5-shot)	46.69%	44.2%	-2.49 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	Exported to ONNX in the ONNX Runtime GenAI graph layout
Quantization	Yes	GPTQ INT4 asymmetric per-group weights (group size 32), dynamic per-token INT8 activations, and an INT8 token-embedding table and LM head; QuaRot R1+R2 rotation applied before quantization; calibrated on 256 wikitext-2 samples
Runtime/backend selection	Yes	ONNX Runtime CPU execution provider with MLAS and KleidiAI
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)
Input type	int64 token ids (input_ids)
Input range	Token ids in the vocabulary range, vocabulary size 128256
Preprocessing	Render chat_template.jinja for the turn, then tokenize the rendered prompt with tokenizer.json, with special tokens present. The template emits the begin-of-text token itself, so the tokenizer is not asked to add it a second time. position_ids is not an input; the ONNX Runtime GenAI graph derives positions internally, and KV-cache I/O is handled by the ONNX Runtime GenAI Generator.

Expected output

Property	Value
Output shape	One token id per generation step, streamed by the ONNX Runtime GenAI Generator
Output type	Token ids
Postprocessing	Decode with tokenizer.json; stop on the eos token or on max length

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.
• 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.

Additional notes

• The ONNX Runtime GenAI artifacts use the fixed names model.onnx (graph) and model.onnx.data (external weights); the runtime requires the .data side-car to sit alongside model.onnx.
• genai_config.json is patched to a max length of 4096, down from the stock 131072, so the KV cache fits the smartphone memory budget; edit it back for longer contexts on higher-RAM devices.
• Model shape: decoder-only Transformer, 16 layers, hidden size 2048, 32 attention heads and 8 KV heads (GQA), vocabulary size 128256, with attention fused into a single GroupQueryAttention operator and the INT4 weights packed into MatMulNBits.
• The k_quant_last recipe leaves the token-embedding table and the final LM head projection at INT8 instead of INT4 — those layers were kept at higher precision to keep accuracy within acceptable ranges.
• MMLU was evaluated with lm-evaluation-harness in the cloze (continuation log-likelihood) format, scoring the four answer letters as continuations of the 5-shot prompt, on a fixed subset of the MMLU test split.

About this version

Original Model: meta-llama/Llama-3.2-1B-Instruct by Meta - 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 Llama 3.2 Community License.

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.