GTC-2.5 & GTC-2.5 mini — The Foundation is Set
400M parameter production-ready foundation model alongside lightweight 64M open-source variant. Validates full training pipeline from data to deployment.
Research
Public research projects, papers, model releases, and technical reports.
GTC-2.5o — Expanding the Horizon
First truly multimodal GTC model. Integrates vision and audio encoders with 0.1B backbone for image captioning, VQA, audio comprehension, and cross-modal reasoning.
GTC-2.5V Preview — Vision-Language Foundation Model
First vision-language model from TAI Research. 64M parameters, visual encoder + decoder-only backbone. Capabilities: image captioning, VQA, text reading.
GTC-2.5 Preview — Conversational Foundation Model
First conversational foundation model. 64M parameters, built on MiniMind. Demonstrates dialogue, knowledge retrieval, code generation.
GTC-Guard-0 — TAI Guardian First Release
Lightweight text safety classifier (<10M params). Built for interpretability and CPU inference. MIT license.
ESFT GTC-2 — Early Supervised Fine-Tuning
Improved conversational performance via early supervised fine-tuning. Weights and repository published.
GTC-2 Large Base Model
Public beta release of the GTC-2 Large base model. Weights and repository published.
GTC-2 Large mini & nano
Lightweight variants with performance close to GTC-2 Large but reduced parameter size.
Sigma Gobang 2 — AI for Gomoku
Enhanced algorithm with multiple difficulty levels. Successor to Sigma Gobang Turbo.
TAI Alignment Lab — Value Alignment & Interpretability
Ongoing research on AI value alignment, mechanistic interpretability, and safety guardrails.