TokForge 1.3.9.4

The current Android beta: Qwen3.5 4B prompts on the NPU, a card that says what your phone gets, and pictures on the GPU of more phones.

Last verified October 20, 2026

Beta since October 20, 2026, on the Google Play beta track and as the Discord APK. Google Play stays on 1.3.6.1 for everyone else. It is the first beta since 1.3.8.1 and includes everything in it; version numbers 1.3.9 to 1.3.9.3 were never released, so their changes are listed here.

• New
• Fixes
• Known issues

New

• Prompts on the NPU for Qwen3.5 4B. On supported Snapdragon phones, the NPU (the phone's AI chip) reads your prompt while the CPU writes the reply. It is opt-in, comes as a one-time pack, and is used only after it passes a check on your phone. More Snapdragon phones can get it now; if your phone does not have enough free memory for it, the check says so and keeps the CPU.
• NPU packs for your phone, a new list in Models > Browse: see which packs fit your phone, and get, stop or remove them.
• What your phone gets, a new card at the top of Settings (Advanced mode) > Speed & engine > Performance > Measured on your phone. It says in plain words what your phone's CPU, GPU and NPU can do for your model, where each check stands and when the next one runs, with a button to run it now.
• A clearer first run. TokForge tells you what your phone can use for chat: the processor, the graphics chip, and the AI chip when there is a pack for it.
• Pictures on the GPU of more phones, used only after a check on your phone shows the same picture, made faster. If the check does not pass, pictures stay on the CPU. You can turn it off in Measured on your phone.
• Faster GGUF replies on some phones, such as the Pixel 9 and the Xiaomi 14T Pro, where the check on your phone picks the GPU. Large GGUF models can get the GPU route from the automatic check too.
• Better Qwen3.5 4B and 9B answers on phones whose processor lacks the i8mm instructions, such as the Snapdragon 865. Models you already downloaded get it too, with no new download.
• The speed add-on reaches more models: Qwen3.5 2B, 27B, 4B Abliterated and 9B Abliterated, used only where your phone measures a gain. If its check fails in the middle of a chat, it switches itself off for the rest of that load.
• Bonsai 2 4B Ternary joins the catalog, and Bonsai models start replying sooner on some phones.
• Qwen3.6 35B-A3B on the CPU shows the thread count it really uses, tells you when a very long message is shortened, and Stop works during a long prompt.
• Terminal packages: jq and sqlite3 ship with the app, each off until you install it from the terminal's Packages screen.
• Model attributions, a new page with each model's credits and licence: Settings > Privacy & about > About > Model attributions.
• Chat > Tune shows the hardware settings before a model is loaded.
• The app is larger than 1.3.8.1, because the NPU runtime files and the terminal packages now ship inside it.

Fixes

• Two possible crashes are prevented: a reply that fails right as it starts while Background Inference is on, and a restart of the Developer API server.
• Qwen3.6 35B-A3B (the MNN build) no longer crashes on long prompts, and no longer risks answering from stale data.
• The speed check waits for a hot phone to cool and then carries on, also on phones that run warm at rest, such as the Pixel 9 and Pixel 10.
• On the Pixel 10, the speed check no longer tries GGUF models on the GPU, where it could stall. They stay on the CPU, and Measured on your phone says why.
• If the graphics driver fails during a GGUF model's first GPU check, the app stays open and the model loads on the CPU.
• Swiping the app away during the background speed check no longer counts as a graphics crash.
• Reload model uses the backend you picked by hand in Chat > Tune.
• A slowdown on Snapdragon 8 Elite Gen 5 phones is gone.
• A model file the app cannot read says so, instead of blaming the GPU. Error messages show a model's name, never its file name.
• Built-in agents run only when you start them, never because a message happens to contain their keyword.
• The download confirmation no longer says you are on mobile data while you are on Wi-Fi with data limits turned on.
• A model set to OpenCL by hand finishes loading (the known issue in 1.3.8.1).
• Importing chats you already have no longer makes copies. Chats doubled by an earlier import stay doubled.
• Model downloads wait for the connection to come back, then continue on their own.
• An occasional wrong reply on the GPU of some phones is fixed.

Known issues

• On phones without i8mm, such as the Snapdragon 865, Qwen3.5 4B and 9B write replies a little slower in exchange for the better answers.
• If you pick the GPU for a model by hand and the phone moves it to the CPU, no message appears yet; the card shows the pick as not used.
• If you start a check with Run now and open a chat while it runs, the chat can show "No model". Tap Load to bring your model back.
• Some German text mixes formal and informal address.

All versions and changelogs · Report a problem