RunAnywhere vs Models Zoo

RunAnywhere brings private, on-device AI to your iPhone and iPad. Download a model once, then chat, transcribe speech, and generate natural-sounding audio locally, even when you are offline. PRIVATE AI IN ACTION Ask questions, summarize ideas, draft text, and have useful conversations without sending every prompt to a remote AI service. VOICE, END TO END - Turn speech into text with compact Whisper models - Listen to natural text-to-speech with Piper voices - Use a voice assistant that connects listening, reasoning, and speaking CHOOSE THE RIGHT MODEL Choose compatible models across llama.cpp, ONNX, and MLX runtimes. Model availability and performance vary by device. BUILT FOR PRIVACY AI inference runs on your device, reducing the need to send prompts or audio to a remote model service. WORK OFFLINE After downloading your selected models, core AI experiences can run without Wi-Fi or cellular data. RunAnywhere is your private AI companion for everyday writing, voice notes, reading aloud, and offline conversations.

Core ML Models is a curated, on-device AI playground. Download any model from the built-in catalog and try it out instantly — no server, no account, nothing leaves your device. Every model is an open-source release converted from PyTorch to Apple's Core ML format and tuned to run efficiently on the Neural Engine, GPU, or CPU. WHAT YOU CAN DO • Chat with small language models (Gemma, etc.) including multimodal vision+text • Generate images from text with Hyper-SD • Remove backgrounds, matte video subjects, and colorize old photos • Estimate depth and reconstruct 3D faces from a single photo • Detect and classify with YOLO, SigLIP, and zero-shot open-vocabulary models • Separate music stems with Demucs and transcribe audio to MIDI with Basic Pitch • Clone voices and synthesize speech with Kokoro and OpenVoice • Super-resolve images with SinSR and RealESRGAN. DEVELOPER TOOLS, NOT JUST A DEMO • Inspect — see every .mlpackage's input/output tensor shapes, data types, compute-unit setting, on-disk size, and Core ML metadata at a glance. • Benchmark — run warmup + timed iterations on your own device and get average / median / min / max latency plus memory delta. Switch compute units (All, CPU+GPU, CPU+ANE, CPU only) to find the fastest config for your hardware. • Swift Code — export a ready-to-paste Core ML loading + inference snippet wired to the model's actual input/output names, so integration into your own app is copy-paste fast. DESIGNED FOR THE NEURAL ENGINE Each model ships with compute-unit settings tuned per-architecture (Neural Engine, GPU, CPU) so you get the best inference speed Apple Silicon can deliver. OPEN AND TRANSPARENT Every model in the catalog links back to its original paper, repository, and license. Conversion scripts are public on GitHub so you can see exactly how the .mlpackage files were produced. PRIVACY FIRST All inference runs locally on your device. The app only fetches model files from public storage over HTTPS — there are no analytics, no ads, and no account to create. REQUIREMENTS • iPhone or iPad with Apple Silicon recommended for best performance • Some large models require several GB of free storage and ≥ 6 GB RAM