OpenThaiGPT 1.6 and R1: Thai-Centric Open Source and Reasoning Large Language Models

TL;DR

This work addresses the challenges of Thai-centric LLMs by proposing two complementary models: OTG-1.6, a 72B generalist trained via Task Arithmetic model merging to integrate diverse Thai-specific knowledge, and OTG-R1, a 32B reasoning-focused model developed through multi-stage training and the Less-Is-More Reasoning (LIMO) approach. Together they demonstrate strong generalization and reasoning capabilities on Thai benchmarks, achieving competitive performance with larger open-source Thai LLMs and even state-of-the-art OpenThaiEval scores. The study provides detailed training regimes, benchmark results, and a discussion of limitations, outlining a practical path toward efficient, high-performing Thai-centric LLMs. Overall, OTG-1.6 and OTG-R1 offer complementary strengths that advance Thai-language reasoning and broad-domain understanding without prohibitive computational costs.

Abstract

We present OpenThaiGPT 1.6 and R1 (OTG-1.6 and OTG-R1), Thai-centric Large Language Models (LLMs) developed through distinct methodologies to enhance generalization and reasoning capabilities. OTG-1.6 employs Task Arithmetic model merging for broad generalization, while OTG-R1 integrates multi-stage training with the Less-Is-More Reasoning Hypothesis (LIMO) for advanced reasoning. Benchmark evaluations demonstrate superior performance across Thai language tasks, achieving competitive results against larger-scale open-source Thai LLMs. This paper details the proposed models, training processes, benchmarks, and results, highlighting improvements over previous models and establishing new performance standards for Thai-centric LLMs.