Frequently Asked Questions

TokForge is a private AI chat app for iPhone and Android that runs large language models (LLMs) entirely on your device. It supports character cards, thinking/reasoning mode, dual inference backends, and built-in benchmarking, all without requiring an internet connection.

No. Once you've downloaded a model, TokForge works 100% offline by default. Internet is only needed to download models from Hugging Face, optionally to check for app updates, and for the new opt-in web search if you choose to turn it on (it's off by default). All AI inference runs locally on your phone's CPU or GPU.

Never. TokForge has zero telemetry, zero analytics, and zero cloud dependency. Your conversations, characters, and settings stay on your device. The optional benchmark sharing feature is entirely opt-in and only shares device specs + performance numbers, never conversation data.

Yes, TokForge is completely free during the beta period. There are no in-app purchases, subscriptions, or ads. Pricing after beta has not been determined yet.

Yes. There's a full native iPhone & iPad build (iOS 16+) at near feature-parity with Android: on-device chat, group chat & roleplay, characters & personas, image generation (Apple CoreML / Neural Engine), vision, voice, RAG, and the agent suite. It's in TestFlight beta: join the iOS beta. Everything runs on-device, same as Android.

TokForge runs on iPhone & iPad (iOS 16+) and any ARM64 Android device running Android 8.0+, which covers the vast majority of phones from the last 6+ years. For a good experience with larger models, you'll want a recent A-series / Snapdragon-class chip and at least 8GB RAM.

We've tested on Samsung Galaxy S20 through S26, RedMagic 11 Pro, OnePlus Ace 5, Xiaomi Pad 7 Pro, and Lenovo Tab Plus, with more devices being added during the beta.

It depends on the model size:

4GB RAM: Qwen3-0.6B runs great (~35+ tok/s)

8GB RAM: Qwen3-4B runs well (~13-20 tok/s depending on GPU)

12GB+ RAM: Qwen3-8B sweet spot (~10-14 tok/s)

16-24GB RAM: Qwen3-14B is possible (~8 tok/s on flagship devices)

TokForge auto-detects your RAM and recommends the right model size during setup.

Model sizes vary by quantization:

Qwen3-0.6B: ~350MB

Qwen3-4B (Q4_K_M): ~2.3GB

Qwen3-8B (Q4_K_M): ~4.8GB

Qwen3-14B (Q4_K_M): ~8.5GB

You can download and delete models freely. TokForge itself is under 50MB.

Yes! The iOS build is universal (iPhone & iPad), and on Android TokForge has a responsive tablet layout, tested on the Xiaomi Pad 7 Pro and Lenovo Tab Plus. Tablets with Snapdragon SoCs get the same GPU acceleration benefits as phones.

TokForge has two inference backends:

MNN (OpenCL GPU): Uses your phone's GPU for inference. Up to 2.4x faster than CPU for models 1.7B+. Best for standard chat and roleplay.

GGUF (llama.cpp CPU): Uses optimized CPU inference with ARM i8mm acceleration. Required for thinking/reasoning mode (<think> blocks). Wins at very small model sizes (0.6B).

TokForge auto-selects the best backend for your device and model combination.

Yes. TokForge includes a built-in Hugging Face model search, so you can find and download any compatible GGUF model directly from the app. There are also 17 curated models across General, Roleplay, Creative, and Thinking categories for easy discovery.

Some models (like Qwen3, DeepSeek-R1, and QwQ) can "think out loud" before answering, showing their chain-of-thought reasoning inside <think> blocks. TokForge renders these as collapsible sections so you can see how the model arrived at its answer. This currently requires the GGUF backend.

TokForge supports the TavernAI V2 character card specification. You can import cards as PNG files (with embedded metadata) or JSON. Features include lorebooks, alternate greetings, world info, {{char}}/{{user}} placeholders, and system prompt assembly.

Yes. TokForge ships with four built-in personas: Rex (direct and analytical), Luna (creative and warm), Marcus (scholarly and thorough), and Aria (artistic and expressive). You can use these as-is or as templates. The v3.6 persona system also lets you build a reusable persona library, bind a persona per conversation, and import personas and whole chat histories from SillyTavern, Layla, and PocketPal.

Yes. TokForge generates images on-device with Stable Diffusion: SD1.5-Turbo, LCM, SD3.x, and an optional NPU Fast tier on rooted devices. You can lock a face or subject across images with reference-image identity (IP-Adapter), apply LoRAs, and run multi-subject and batch generation. No cloud, no API fees, and it works in airplane mode.

Yes. Attach a photo and ask about it. Qwen3-VL and Gemma-4 Omni run vision natively, and SmolVLM ships as a sidecar for models without built-in vision. Gemma-4 Omni also handles text, audio, and video, all processed on-device.

Yes. Build a reusable persona library and bind one per conversation, then use the dedicated roleplay mode with lorebooks, multi-character group chat, and rotation. You can import characters and whole chat histories from SillyTavern, Layla, and PocketPal.

Optionally. Web search in chat is opt-in and off by default, and it's consent-gated each time. TokForge stays fully offline unless you explicitly turn it on.

Agents let you build no-code, multi-step workflows that call tools and run entirely on your device, with bounded loops and triggers. TokForge ships with built-in agents and a simple builder, including a batch image-generation agent.

Tokens per second (tok/s) measures how fast the model generates text. A token is roughly 3/4 of a word. For comfortable reading speed, you want at least 5-8 tok/s. Above 15 tok/s feels nearly instant. TokForge shows tok/s in real-time during conversations and in ForgeLab benchmarks.

ForgeLab is TokForge's built-in benchmarking suite. It measures decode tok/s, prefill latency, and throughput across every model and backend on your device. The Auto-Matrix feature tests all combinations automatically and saves results in a persistent local database. You can export results and share them with the community.

On Android, TokForge is in open testing on the Google Play Store. On iPhone & iPad it's in TestFlight beta: join here. You can also grab either from the homepage. We prioritize testers with diverse devices to help build our cross-device benchmark database.

We're working toward a public release on the Google Play Store. The timeline depends on feedback from the current open beta. Join the beta to help us get there faster and shape the final product.