TokForge Local AI Offline Chat, from Defcon-One, puts large language models on Android for private, offline interaction and content work. The app performs full inference locally on CPU, GPU (Adreno/Vulkan), or NPU and supports GGUF and MNN model formats for chat, image generation, roleplay, and document Q&A. It targets privacy-conscious users, travelers, and researchers who need AI access without cloud services, offering local model execution, voice interaction, and file analysis for mobile routines.

Runs full models on the device to keep conversations and files local

TokForge executes Large Language Models entirely on the handset, so prompts and document queries do not leave the phone. It supports running on CPU, GPU (Adreno with Vulkan), or NPU, and accepts GGUF and MNN formats, enabling chat, on-device image generation, and speech features without external servers. This design enforces local-only inference and suits users who must keep transcripts, lorebooks, and processed files on-device.

Adapts processing to available hardware but needs ample memory for large models

The built-in AutoForge benchmark profiles the device to choose the fastest backend, routing work to CPU, GPU, or NPU as appropriate. The app recommends devices with 8GB+ RAM for larger models, and it notes GPU/NPU acceleration on compatible Snapdragon and Adreno chips. Expect performance to scale with RAM and SoC capabilities, since heavier models demand more memory and compute resources.

Supports offline document Q&A and character-driven roleplay workflows

For practical daily use, the app parses local PDFs, DOCX, EPUB, Markdown, CSV, and text files so users can ask questions about their documents without sending them online. Character roleplay imports PNG/JSON cards from chub.ai and TavernAI, and lorebooks plus alternate greetings let writers iterate locally. These tools target students, writers, and travelers who craft content away from network access.

Enforces a clear privacy posture but poses advanced configuration choices

TokForge ships with zero analytics and zero telemetry, requires no account, and stores conversations and assets on-device, reflecting a privacy-first stance. Voice cloning, text-to-speech, and speech-to-text operate alongside the offline model capability, enabling voice workflows in airplane mode. The developer also exposes a large local API surface (over 270 endpoints), which offers automation power but adds complexity for casual users. Community feedback highlights strong privacy and notes performance varies by device.

To sum up, the app suits privacy-minded mobile users who need offline AI

To sum up, the app is a practical option for privacy-minded users who need reliable AI access during travel or in offline environments; its usefulness falls if your device has limited RAM or lacks hardware acceleration. As a mobile, local-first tool, it fits people who prioritize data sovereignty and are comfortable tuning backend settings for better performance.