Our Models
Trained on Traditional Chinese linguistic context, these models bring reasoning and dialogue closer to the way we actually live and speak
T1 Model
A high-speed favorite based on Gemma-3-4B. Deep-rooted Taiwan knowledge makes it the go-to for natural, precise local dialogue.
Formosa-1
Logic-first model grounded in Llama-3.2-3B. Engineered for deep reasoning and robust answers to complex prompts.
Soon
Twinkle Voice
A voice-first branch. Because sometimes, your hands deserve a break from typing.
Evaluation System
To quantify how deeply a model understands Taiwan, we built a specialized evaluation tool that measures cultural nuances.
Twinkle Eval Leaderboard| Name | Model Type | Parameters | Tags | |
|---|---|---|---|---|
T1 Series | ||||
| gemma-3-4B-T1-it | text-generation | 4B | gemma3_texttext-generationTaiwanR.O.C | 41 |
| gemma-3-4B-T1-it-GGUF | text-generation | 4B | TaiwanSLMagenttext-generation | 4 |
| gemma-4-T1-E4B-it | image-text-to-text | 4B | image-text-to-textenzhbase_model:google/gemma-4-E4B | 0 |
Formosa-1 Series | ||||
| Llama-3.2-3B-F1-Reasoning-Instruct | text-generation | 3B | llamatext-generationTaiwanR.O.C | 48 |
| Llama-3.2-3B-F1-Instruct | text-generation | 3B | llamatext-generationTaiwanR.O.C | 27 |
| Llama-3.2-3B-F1-Reasoning-Instruct-GGUF | text-generation | 3B | TaiwanR.O.CzhtwSLM | 14 |
Other Models | ||||
| twinkle-sqlcoder | text-generation | N/A | mistraltext-generationtext-to-sqlsql | 2 |
| TwinkleTokenizer | N/A | N/A | tokenizersTaiwanR.O.Czhtw | 0 |