core — shared files every other group needs
inference — generate or score on device
train — fine-tune on device, then merge the adapter back into the base weights

Fine-tuning method

mars — MARS (Multi-Adapter Rank Sharing) — this project's own method: adapters shared across layers, so parameter count grows with rank rather than with depth.
Rank: 8
Adapted modules: q_proj, v_proj

Provenance

Base model: google/gemma-3-270m-it
Selected task: text-generation-with-past
Quantization: int4
Toolchain: optimum-onnx 0.1.0, transformers 4.57.6, ort-training 1.23.0+cpu

Licenses

Framework: not declared in this package — see the repository
Base model weights: see the base model above (this package redistributes an export of those weights, so their terms govern its contents)

Android runtime

Minimum API: 28
Required ABIs: any

Variants

id	EP	quant	engines	features	min API	rec. RAM (MB)
cpu-int4	cpu	int4	native	core, inference, train	28	—

Default variant: cpu-int4.

Running this model

This is a MobileTransformers package, not a plain Hugging Face model: it is a manifest plus per-variant ONNX stages and a weight-handoff map. transformers, optimum and plain onnxruntime cannot load it. Use the framework:

https://github.com/martinkorelic/mobiletransformers

kotlin

// Android — pulls, verifies and installs on first use.
val model = MobileTransformers.fromPretrained(
    context = context,
    repoId  = "mobiletransformers/gemma-3-270m-it",
)

bash

# Host — download and inspect the package without a device.
mobiletransformers pull --repo-id mobiletransformers/gemma-3-270m-it

Citation

If you are using this framework for your own work, please cite:

bibtex

@misc{mobiletransformers2025,
  author       = {Koreli\v{c}, Martin and Pejovi{\'c}, Veljko},
  title        = {MobileTransformers: An On-Device LLM PEFT Framework for Fine-Tuning and Inference},
  year         = {2025},
  howpublished = {\url{https://gitlab.fri.uni-lj.si/lrk/mobiletransformers}}
}