trio-input v1.0.1 — Windows installer

Download TrioInput-Setup-1.0.1.exe and run it. No git, no Python required.

What you get: a real Windows input method, Trio Input 中/EN/SV, in the language bar. Switch to it with Win+Space and type Chinese pinyin, English and Swedish in any application without switching again. The setup detects your GPU and pre-selects the matching runtime (NVIDIA → CUDA ~4.5 GB, Intel Arc / Core Ultra iGPU → XPU ~3.2 GB, otherwise CPU ~1.7 GB), downloads the Python runtime and the Qwen3-0.6B model from this page, installs the PIME text service (one administrator prompt) and registers the input method. A local page at http://127.0.0.1:8766 shows probabilities, learning status and the retrain button.

Unattended: TrioInput-Setup-1.0.1.exe /VERYSILENT /GPU=auto|cpu|xpu|cuda. Uninstall from Windows Settings → Apps (removes the input method too).

Features: n-gram + local-LLM language ranking, always-visible per-language candidates, ä/ö/å via the US-layout keys (' ; [), spelling correction, predict-ahead (Tab), learning from your selections with one-click retrain.

Other assets on this page are downloaded by the installer: python-runtime-<cpu|xpu|cuda>-win64*.zip (embeddable Python 3.12 + PyTorch), the Qwen3-0.6B model files (Apache-2.0) and a mirror of PIME-1.3.0-stable-setup.exe (EasyIME/PIME, GPL).

trio-input v1.0.0

Final release of the standalone prototype: Chinese pinyin / English / Swedish in one input, no switching.

Install (Windows x64): clone or download the source, then run powershell -ExecutionPolicy Bypass -File install.ps1. The installer downloads the model assets below as needed (Qwen3-0.6B-int8.pt + tokenizer for CPU; model.safetensors additionally for NVIDIA / Intel GPU). Device order at runtime: NVIDIA CUDA -> Intel XPU -> CPU.

Features: n-gram + local-LLM language ranking with always-visible per-language escape hatches; common-word priors for all three languages; ä/ö/å via the US-layout keys; spelling correction; predict-ahead with Tab; learning from selections with one-click retrain. (Command-line mode was removed before this release.)

Assets: Qwen3-0.6B (Apache-2.0) — Qwen3-0.6B-tokenizer.zip, Qwen3-0.6B-int8.pt (~1.2 GB, CPU), model.safetensors (~1.5 GB, GPU).

trio-input v0.1.0

First release of the standalone prototype: pinyin / English / Swedish in one input, no switching.

Install (Windows x64): clone or download the source, then run powershell -ExecutionPolicy Bypass -File install.ps1. The installer downloads the model assets below as needed (Qwen3-0.6B-int8.pt + tokenizer for CPU; model.safetensors additionally for NVIDIA / Intel GPU).

Assets: Qwen3-0.6B (Apache-2.0) — Qwen3-0.6B-tokenizer.zip (tokenizer + config), Qwen3-0.6B-int8.pt (dynamic-int8 CPU model, ~1.2 GB), model.safetensors (bf16 weights for GPU, ~1.5 GB).

Features: n-gram + local-LLM language ranking, always-visible escape hatches per language, ä/ö/å via the US layout keys, spelling correction, predict-ahead with Tab, learning from selections with one-click retrain.