ayushg-20/ondevicellm-benchmark-converted-models

OnDeviceLLMBench Converted Models

Runtime-specific converted model artifacts for the OnDeviceLLMBench iOS benchmark app.

This repository intentionally contains only artifacts that were not already available in the required runtime layout on Hugging Face at the time of upload. Models that already have suitable public artifacts are referenced directly by the app scripts instead of mirrored here.

Contents

Path	Base model	Runtime format	Notes
Qwen3-0.6B-coreml-int4.mlmodelc/	Qwen/Qwen3-0.6B	Core ML compiled model	Local int4 Core ML conversion for EngineCoreML.
Qwen3-0.6B-CoreML-LLM/	Qwen/Qwen3-0.6B	CoreML-LLM bundle	john-rocky CoreML-LLM layout: model.mlpackage, model_config.json, hf_model/.
Qwen3.5-0.8B-onnx-cpu-int4/	Qwen3.5 0.8B	ONNX Runtime GenAI	Single-folder GenAI layout with genai_config.json, model.onnx, and external data.
SmolLM2-135M-MLX-8bit/	HuggingFaceTB/SmolLM2-135M	MLX 8-bit	Base-model MLX 8-bit artifact used by the benchmark.
SmolLM2-135M-litert-dynamic-wi8-afp32.litertlm	HuggingFaceTB/SmolLM2-135M	LiteRT-LM	Dynamic int8-weight / fp32-activation LiteRT-LM bundle.
SmolLM2-135M-executorch-int8/	HuggingFaceTB/SmolLM2-135M	ExecuTorch	XNNPACK int8 .pte export plus tokenizer/config.
SmolLM2-135M-onnx-cpu-int8/	HuggingFaceTB/SmolLM2-135M	ONNX Runtime GenAI	CPU dynamic int8 GenAI layout.

Provenance

Conversion scripts live in the OnDeviceLLMBench repository under scripts/. Upstream model licenses and use restrictions remain applicable. Qwen and SmolLM2 upstream checkpoints are Apache-2.0.

Qwen2.5 artifacts are intentionally excluded from this repository and from the benchmark manifest.