TER-AI
TER-AI is an independent, AI-assisted applied research and development project for Thai-language AI systems, covering a typo-tolerant Thai intent router, local CPU-first chat and vision, Thai speech, SVG-based graphics, and audio and music tools[^c1]. It is distributed as the Python package terai-router, a small Thai-language toolkit designed to run fully offline on CPU[^c2]. The project combines components built specifically for it with existing open models, and it states explicitly that it does not train its own foundation model[^c3].
The offline-first design is the project's defining characteristic. The intent router, the rule-based chat, the TERM AI legacy chatbot, and the SANKPUP offline image generator all run on the CPU without a network connection[^c4]. The router reads one Thai or Thai-English message, repairs common typographical and speech-to-text errors, and returns one of fifteen intents with a confidence score from a 1.1 MB dependency-free model[^c5]. Chat answers with deterministic tools and a bundled Thai Wikipedia pack of 8,192 articles[^c6], while optional web search extends it to Wikipedia and a user-supplied SearXNG instance. TTH provides Thai speech synthesis[^c7], SANKPUP generates pictures by writing and sanitizing SVG rather than by running an image model[^c8], and the MuML systems compose music both in the browser and as rendered audio files[^c9]. Since v0.16.0, the optional TERAI1 layer uses a local Gemma model to write chat replies from verified tool results[^c10].
The project pairs its capabilities with an unusual emphasis on reproducibility. Frozen test sets, published dataset hashes, a local router scorecard[^c11], and JavaScript-to-Python parity checks are part of the repository. Package releases and model identifiers version independently[^c12], and licensing is split by component: the code is MIT, while some bundled data and optional models carry separate terms[^c13]. The v0.18.0-beta1 release is a browser-demo and documentation update rather than a new model claim[^c16].
TER-AI remains an experimental, single-maintainer project. The repository was created on 2026-10-16[^c15], and its maintainer describes it as an indie effort that may pause or close with little warning[^c14]. Its documentation, evaluation tooling, and licensing are therefore best read as a transparent snapshot of a rapidly evolving project rather than a guarantee of long-term support.