Home / Features
What your phone gets
TokForge does not guess. It checks what your own phone does well, times it, and uses the faster way, so two phones with the same model can end up with different setups.
Last verified October 13, 2026
On every phone
- Private chat, offline. Once a model is downloaded, it answers in airplane mode. No account, nothing sent to a cloud model.
- Characters. Four included, build your own, group chats, and character cards from SillyTavern and others.
- Pictures. Image models run on the phone, with no per-image fee.
- Your files. Ask about a PDF, Word file or ebook; answers point back to the file.
- Agents. 12 built-in agents. From 1.3.8 the Agents tab has a run screen with live steps and three starter presets.
- Voice. Voice typing and replies read out loud.
- 8 app languages on Android, and a local control API for developers.
Your phone picks the route 1.3.8 and newer
- A check, then a timing. While the app sits idle, TokForge compares the GPU's answers with the CPU's for the model you loaded. If they agree, it times the CPU, the GPU and a mix of both (GPU reads the prompt and CPU writes the reply, or the other way round) and keeps whichever measured faster on your phone. If the check fails, the model stays on the CPU and its card says why.
- You can see it and change it. Settings (Advanced mode) > Speed & engine > Performance > Measured on your phone shows what it found, with Off, Auto and On for each route and speed feature, and one switch that covers all of them.
- It stays out of your way. Checks run only while the app is idle in the foreground, the phone is cool, the battery is not low and power saving is off, and they stop when you start typing.
- It checks again when something changes: a new app version, a system update, a new GPU driver or a new model file.
Phone by phone
What the current beta (1.3.8.1) does on the phones we test, and on everything else. Google Play 1.3.6.1 does not have these measured routes yet.
| Phone or chip | What TokForge does |
|---|---|
| Snapdragon 8 Gen 3, Snapdragon 8 Elite Gen 5 | CPU, GPU and mixed routes, measured on your phone. An optional NPU speed boost reads long prompts for Qwen3 0.6B, 1.7B and 4B Abliterated (opt-in, checked on its first load). |
| Pixel 9 | Measured routes, including the "GPU prompt, CPU reply" mix for MNN models. Long prompts start sooner on the CPU, and GGUF replies use four CPU threads. |
| Xiaomi 14T Pro | Long prompts start sooner on the CPU. MNN on OpenCL stays off, and GPU setups respect the system's per-app memory cap. |
| OnePlus Ace 5 Ultra | Measured routes. The GPU check runs inside the app instead of in a separate helper, and OpenCL works normally. |
| Pixel 10 | Stays on the CPU for now, even when you pick Vulkan yourself. |
| Snapdragon 865 | Stays on the CPU for now. GGUF on Vulkan is not offered. |
| Any other phone | Measured too. GGUF on Vulkan reads your phone's real GPU name and gets crash protection, and a GPU route is used only where its check passes and it measures faster. |
Speed add-ons 1.3.8.1 and newer
- Multi-token prediction is an optional add-on for Qwen3.5 4B and 9B, used only where your phone measures a gain.
- It rides along with the model download. If you already have the model, tap Get on it. Automatic add-on downloads wait for Wi-Fi and ask first.
- Prompt lookup (for MNN models) and the Llama draft pack are measured the same way and turn on only where your phone shows a gain.
NPU, the phone's AI chip
- Speed boost 1.3.8 and newer: on Snapdragon 8 Gen 3 and Snapdragon 8 Elite Gen 5, an optional download lets the NPU read your prompt for Qwen3 0.6B, 1.7B and 4B Abliterated. It stays off until you turn it on, checks itself on its first load, and stays on the CPU if that check fails.
Memory
- Models are sorted into Best for your phone, Fits and Too big, using the same memory question as loading.
- For GGUF models, the "CPU prompt, GPU reply" mix can share one copy of the model between CPU and GPU instead of two, where your phone measures it worth it.
- On phones whose system caps how much memory one app may use, GPU setups and speed checks respect that cap.
How fast will it be?
It depends on your phone, its memory and the model you pick, so we do not quote one number for everyone. The figure that counts is the one your own phone measures: open Settings (Advanced mode) > Speed & engine > Performance > Measured on your phone. To compare with other people's phones, see the community speed leaderboard, where every result was submitted from the app on a real device.
Next: the models · download · how-to