Local LLM runner vs Abliterated

Run LLM on your iPhone, 100% local, no network (Network needed once to download a model) Models: • Qwen2.5 0.5B Instruct — ~0.4 GB (4 GB RAM) • Llama 3.2 1B Instruct — ~0.8 GB (4 GB RAM) • Qwen2.5 1.5B Instruct — ~0.9 GB (4 GB RAM) • Llama 3.2 3B Instruct — ~1.9 GB (6 GB RAM) • Qwen3 4B Instruct 2507 — ~2.5 GB (6 GB RAM) • Qwen2.5 7B Instruct — ~4.4 GB (8 GB RAM) Also: • Add your own GGUF (for example, from Hugging Face) • Auto-estimated max context from GPU memory; you can change it Requires iOS 18 or later. Tested on iPhone 11, 14 Pro, 15 Pro with iOS 26. Qwen models © Alibaba Cloud. Licensed under Apache License 2.0. Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. Built with Llama. On-device inference uses llama.cpp (MIT License).

Abliterated is an unofficial native iOS client for models hosted by abliteration.ai. Talk to abliterated models: open-weight LLMs with the refusal direction removed in the weights. They answer prompts other APIs decline. Use it for writing, code, research, red-teaming, and eval work the hosted models are built for. You are responsible for how you use the models and for following Apple’s and local law. Disclaimer: You need an abliteration.ai account and API key. Usage is billed by abliteration.ai. No affiliation with Abliteration AI. Conversations go to abliteration.ai. Abliteration states zero data retention by default.