GTC Series
GTC-2.5 & GTC-2.5 mini
TAI Research's most capable foundation model to date. GTC-2.5 (400M parameters) delivers substantially improved coherence, knowledge retention, and instruction-following capabilities. GTC-2.5 mini (64M) is open-sourced for the community.
- 19GB curated dataset — Chinese, English, code, mathematics
- Significantly better factual accuracy and reasoning
- Stable multi-turn dialogue and reliable instruction following
- GTC-2.5 mini available under permissive license
GTC-2.5o
The first truly multimodal member of the GTC family. Integrates vision and audio encoders directly into the foundation, enabling unified understanding across images, speech, and text.
- Image captioning, VQA, and visual scene interpretation
- Audio comprehension and spoken input processing
- Cross-modal reasoning across multiple sources
- Baseline for larger multimodal models
GTC-2.5 Preview
TAI Research's first conversational foundation model. Built on the open‑source MiniMind project, GTC‑2.5 Preview validates our training pipeline and serves as a baseline for larger models to come.
- Dialogue, knowledge retrieval, and code generation
- Compact 64M design — fast inference, low footprint
- Designed for technical validation and internal benchmarking
GTC-2.5V Preview
TAI Research's first vision‑language model. Extending GTC with a lightweight visual encoder, GTC‑2.5V Preview supports image captioning, VQA, and OCR‑style text reading from images.
- Multimodal understanding — images + text
- 64M backbone + visual encoder
- Baseline for larger multimodal models