TokForge Local AI Offline Chat

Defcon-One

Bu ilova haqida

TokForge 1.0 is here. Full AI models run directly on your phone. No cloud, no account, no subscription. Your conversations never leave your device.

Chat and roleplay. Create characters or import cards from chub.ai and TavernAI (PNG/JSON). Personas, lorebooks, alternate greetings, and group chats where every character can speak with its own voice.

Generate images on-device. Type a prompt and get a picture without an internet connection. Runs on CPU everywhere, with GPU acceleration on Adreno and NPU acceleration on supported Snapdragon devices.

Talk to it. Kokoro text-to-speech with multiple voices, plus voice cloning: record about a minute of audio and TokForge speaks in that voice. Voice input included.

Ask your documents. Drop in a PDF, DOCX, EPUB, Markdown, CSV, or text file and ask questions about it. Long-term memory and a knowledge graph keep track of what matters across conversations.

Web search, when you want it. Off by default. Turn it on to give the model current information.

Three engines, tuned for your chip. llama.cpp (GGUF) with OpenCL and Vulkan GPU paths, MNN with OpenCL, or connect to your own OpenAI-compatible server. SoC-aware routing picks the fastest path automatically.

Find and benchmark models. A curated catalog of 52 models matched to your device, plus a built-in downloader with Hugging Face search. AutoForge benchmarks every model and backend combo on your hardware and posts to the public leaderboard at leaderboard.tokforge.ai if you choose.

Backup and restore your conversations, characters and settings. Developer API with 270+ local endpoints. Zero analytics, zero telemetry, works in airplane mode. Available in 8 languages.

Community.
Discord: discord.gg/Acv3CBtfVm
Leaderboard: leaderboard.tokforge.ai
Web: tokforge.ai

iPhone and iPad: TokForge is available as a public beta on TestFlight.

Nima yangiliklar

Version 1.2.1. Responsiveness fixes: the app no longer competes with itself for CPU while a model loads, so the screen stays smooth during loading. Image generation gains extra protection against rare freezes on some GPUs. Several background storage checks were moved off the main thread. The typing indicator now uses less power, and replies start slightly faster.