noni - AI Assistant

noni is a team of small AI models that live on your device. Instead of one giant model in the cloud, noni orchestrates several small open-weight language models locally: a tiny router reads each request and hands it to the right specialist - general chat, coding, or writing. Nothing you type ever leaves your device. PRIVATE BY ARCHITECTURE • zero data collection - no account, no analytics, no tracking • conversations are stored only on your device • the network is used for one thing: downloading model weights from Hugging Face • after a model is downloaded, it works in airplane mode A TEAM, NOT A SINGLE MODEL • auto - a 0.6B router classifies each request and picks the specialist • solo - pin any downloaded model and talk to it directly • relay - pipelines like draft → polish, one model refining another's work • council - two models answer, a third merges the best of both BUILT TO RUN ON YOUR DEVICE • Apple's MLX framework with 4-bit quantized models and streaming replies • markdown rendering with copyable code cards • background, resumable model downloads • automatic conversation titles, edit & resend, regenerate • add any public model from Hugging Face by pasting a link • a strict black-and-white design: no clutter, no color, no distractions MODELS & SIZES (download only what you need) • qwen3 0.6b - router and quick chat, ~350 MB • llama 3.2 1b - everyday general model, ~710 MB • gemma 3 1b - rewriting and summarizing, ~770 MB • qwen2.5 coder 1.5b - writes and explains code, ~880 MB • qwen3 1.7b - stronger generalist for recent devices, ~980 MB • llama 3.2 3b - the most capable, for high-memory devices, ~1.8 GB The starter pack (router + general model) is about 1.1 GB. Models are free, open-weight, and hosted on Hugging Face. HONEST EXPECTATIONS Small local models are fast, private, and surprisingly capable for everyday questions, quick code snippets, and rewrites - but they are not cloud giants. Expect roughly 10-30 tokens per second on recent devices, and occasional mistakes. That is the trade for total privacy, zero cost, and offline use. Requires iOS 17 or later. On older or lower-memory devices, stick to the smaller models; the library labels which models want a high-memory device. Built with Llama. Qwen models are Apache-2.0 by Alibaba Cloud; Gemma is provided under Google's Gemma Terms of Use; inference runs on Apple's open-source MLX.