Features

Model discovery: search Hugging Face Hub for GGUF models with hardware-compatibility badges
Download manager: background downloads with pause/resume and queue management
Local chat: streaming inference with markdown rendering, system prompts, and conversation history
Model configuration: temperature, top_p, top_k, context length, JSON mode, and more
Multi-model library: manage installed models and import from the Files app
Local API server: OpenAI-compatible endpoints (/v1/chat/completions, /v1/models, /v1/embeddings)
Multi-modal: vision-model support with image inputs
Chat templates: auto-detects prompt format (ChatML, Llama, Mistral, Phi, Gemma, and more)
Speculative decoding: draft-model acceleration for faster inference
Performance dashboard: real-time tokens/sec, memory usage, and thermal monitoring

Requirements

Component	Technology
Language	Swift (strict mode)
UI	SwiftUI with iOS 26 liquid glass
Inference	llama.cpp (Metal GPU acceleration)
Model source	Hugging Face Hub API
Persistence	SwiftData
Local server	Swift NIO
Architecture	MVVM with Clean Architecture

All inference runs on-device. No cloud backend. No analytics. No tracking. Chat history and models are stored locally only.

Frequently Asked Questions

Does LM Pocket send any data to the cloud?

No. All inference runs on-device with no cloud backend, no analytics, and no tracking; chat history and models are stored locally only.

What models can I run?

Any GGUF-format model discoverable through the Hugging Face Hub, with hardware-compatibility badges to indicate what will run well on your device.