Which AI models run on your phone

TokForge comes with a curated list of models, each tagged with how much phone memory it needs. The app shows you only what fits, sorted into Best for your phone, Fits and Too big. Every download comes from a public Hugging Face repository, linked below.

Last verified October 20, 2026

• Chat models
• Picture models
• Helpers
• Speed add-ons
• NPU packs
• Model guides
• Your own models
• By version

The list has 53 curated downloads on Android: 34 chat downloads (28 models, some in both an MNN and a GGUF build), 18 picture models (15 offered for new downloads) and a vision helper. The 1.3.9.4 beta adds Bonsai 2 4B Ternary to the list Google Play 1.3.6.1 has; what else changed is where some of them download from and the add-ons. The iPhone and iPad beta has its own list.

Chat models

Sorted by the phone memory each one needs. "Phone memory" is the app's own gate: below it, the model is marked Too big instead of failing halfway through loading. Abliterated models answer more freely than their originals.

Picture models

Shown in the Image tab when your phone passes that model's check. NPU models use the AI chip of supported Snapdragon phones.

Helpers

• SmolVLM 500M (Vision Sidecar): describes the pictures you attach, for models that cannot see. Hugging Face
• Document helpers (two small embedders and a summarizer) for questions about your files. TokForge offers them once, shows their size, and downloads nothing if you choose Later (1.3.8.1 and newer).
• Voice: English voices to read replies out loud, optional voice cloning, and speech recognition that runs on the phone.
• Acceleration packs: small draft models used by speculative decoding, downloaded only with a model they pair with.

Speed add-ons

An optional multi-token prediction add-on for some Qwen3.5 models, used only where your phone measures a gain. It downloads with the model; if you already have the model, tap Get on it. How it works.

NPU packs

Opt-in packs that let the NPU of supported Snapdragon phones read your prompt while the CPU writes the reply. Each turns on only after it passes a check on your phone. More.

All packs are in the TokForge NPU packs collection. From 1.3.9.4, Models > Browse lists the packs that fit your phone's chip, which can include packs from the collection that are not in the table above.

Model guides

• Qwen3.5 0.8B: the smallest model, for phones with 4 GB.
• Qwen3.5 2B: for phones with 6 GB.
• Qwen3.5 4B: the size most people settle on, for 8 GB.
• Qwen3.5 9B: for phones with 12 GB.
• Gemma 4 E2B: for 8 GB.
• Can my phone run it? Check a model against your phone before you download.
• Model size guide and MNN or GGUF.

Your own models

Beyond the curated list, you can search Hugging Face from the app (Browse) and add GGUF files yourself. TokForge checks them against your phone's memory the same way.

By version

• 1.3.9.4 (beta): Bonsai 2 4B Ternary joins the list, Qwen3.5 2B and 27B move to the mirrors with add-ons, and the Qwen3.5 4B NPU pack arrives.
• 1.3.8.1: Qwen3.5 4B and 9B download from mirrors that carry the speed add-on, Qwen3 1.7B downloads a smaller build, and model names say Abliterated.
• 1.3.6.1 (Google Play): 52 downloads, the same list without Bonsai 2 4B Ternary and without the speed add-ons.