TokForge 1.3.9.7

The current Google Play version: LiquidAI LFM3 models, a faster start for Qwen3.5 9B on the NPU, replies that start sooner, and many memory and stability fixes.

Last verified October 20, 2026

On Google Play since November 6, 2026, and on the Google Play beta track and as the Discord APK. It includes the 1.3.9.4 beta and the 1.3.9.5 and 1.3.9.6 betas, so their changes are listed here.

• New
• Fixes
• Known limits

New

• LiquidAI LFM3 models. LFM3 text models run in TokForge, in GGUF and, from 1.3.9.7, as a smaller MNN build that uses less memory. TokForge treats LFM3 as a thinking model, so the Thinking switch applies, and LFM3 can call tools in the Agents tab. It runs on the processor and, on phones with a supported GPU such as the Pixel 9 and the Xiaomi 14T Pro, on the GPU through Vulkan.
• Replies start sooner. When you open the app, TokForge gets the model you used last ready, and a new chat reads its opening instructions while you type, so the first words arrive sooner. A switch in Settings > Speed & engine > Performance, "Get my last model ready when the app opens", is on by default.
• Reopened chats pick up where they left off, and your next message starts sooner.
• Chats keep their memory between turns. Qwen3.5 and Qwen3.6 chats, on both MNN and GGUF, read only your new message on a follow-up instead of the whole chat again, so replies stay quick as a chat grows.
• Qwen3.5 9B starts much faster on the NPU, and the speed add-on now works together with NPU prompt reading on Qwen3.5 4B.
• Brazilian Portuguese is now a full app language, and the starter characters speak your app language in every language TokForge offers, with their names unchanged.
• New models. MiniCPM-V 4.7 1B joins the MNN catalog (and MiniCPM-V 4.7 35B-A3B for phones with 24 GB), Bonsai 2 4B and Bonsai 2 9B join as MNN models, and Bonsai 2 4B Ternary is in the list. The speed add-on now also covers Qwen3.5 0.8B, used only where your phone measures a gain.
• A quieter first speed check. It waits until your first chat has finished, or until the phone is idle, cool and charging or well charged, and never runs while you type or a reply is being written. "Run the check now" still starts it right away.
• The screen stays responsive while a model loads or writes a reply.
• Thinking models answer once their thinking budget is used up, on both MNN and GGUF, instead of carrying on reasoning in the reply.
• Safer backups. Restoring a backup is done across a restart: TokForge prepares the backup, restarts itself and applies it before anything opens your data, and if it cannot finish, your previous data is kept.

Fixes

• The background speed check no longer crashes some Snapdragon 8 Elite phones.
• Each chat keeps its own saved state, and the dates on your saved memory facts are right.
• Many more checks on model files: a damaged or malformed word list, weights file, model graph or chat template, in MNN or GGUF models, is refused with a reason instead of crashing the app. A word list in the newest format now loads.
• Switching models when both do not fit in free memory unloads the old model first and waits for its memory to come free, so fewer of your other apps have to restart.
• The first GPU check of a GGUF model starts only when the phone has enough free memory; if memory is short, the model runs on the processor and the check waits for a quieter moment. On 8 GB phones the check no longer makes Android close your other apps.
• If the graphics driver fails during a GGUF model's first GPU check, the check ends quietly: the model runs on the processor, and the driver fault is no longer reported as a TokForge crash.
• Qwen3.5 4B and 9B Abliterated give better answers on phones without i8mm, such as the Snapdragon 865, as the standard Qwen3.5 models already did.
• Gemma 4 GGUF models chat in their own format, so replies no longer end with stray tag text.
• In the Agents tab, tool results and the agent's instructions reach models whose chat format has no slot for them, the same way they already did in chat.
• Read aloud reads long replies in full and no longer holds up the screen while the voice engine starts.

Known limits

• On phones without i8mm, such as the Snapdragon 865, Qwen3.5 4B and 9B (Abliterated included) write replies a little slower in exchange for the better answers.
• On some phones, such as the Pixel 10, GGUF models run on the processor: the app keeps them there for that graphics family.
• LFM3 runs as a text model; LFM3 vision models (image input) are not supported yet.
• The terminal is a first version: it stays off until you turn it on, and it is available only where your phone passes its check.

All versions and changelogs · Report a problem