Today, we are releasing GTC-2.5 — our most capable, production-ready foundation model to date — alongside GTC-2.5 mini, a lightweight, open-source variant for the community.

GTC-2.5: A 400M Parameter Leap

GTC-2.5 is the first model in the GTC series that we consider truly usable. At 400 million parameters, it represents a significant scale-up from our earlier 64M Preview models. Trained on a carefully curated 19GB dataset — combining high-quality Chinese and English text, code, and mathematical content — GTC-2.5 delivers substantially improved coherence, knowledge retention, and instruction-following capabilities.

The model was trained on V100 GPUs and fine-tuned through a rigorous SFT process. While 400M is modest by industry standards, it is a meaningful milestone for us: it validates our entire training pipeline, from data collection to model deployment, and provides a solid foundation for everything we plan to build next.

Key improvements over GTC-2.5 Preview:

  • Significantly better factual accuracy and reasoning
  • More stable multi-turn dialogue
  • Improved code generation (still limited, but notably more useful)
  • Reliable instruction following in both Chinese and English

GTC-2.5 is now available on Sigma and through our API for Pro users.

GTC-2.5 mini: Open for the Community

Alongside GTC-2.5, we are open-sourcing GTC-2.5 mini, a 64M-parameter model derived from our earlier Preview work. It is lightweight, fast, and easy to deploy — suitable for research, experimentation, and lightweight edge applications.

Why open-source a smaller model? We believe in making AI accessible. GTC-2.5 mini is not our most powerful model, but it is a solid starting point for anyone who wants to experiment with a compact, chat-capable foundation model without heavy infrastructure requirements. It is also fully compatible with our broader GTC ecosystem, meaning you can easily upgrade to larger models as your needs grow.

GTC-2.5 mini is available now under a permissive license on Hugging Face and GitHub.

What This Means for TAI Research

With GTC-2.5, we now have a stable, usable foundation model that can support real-world applications — from Sigma and TAI Teens to enterprise partnerships. The model is not perfect, but it is a functional baseline that we can iterate on, scale up, and adapt for specialized tasks.

This release marks the end of the "Preview" era for GTC and the beginning of a more mature, product-oriented phase. We are no longer just experimenting — we are building.

Next up: GTC-2.5 Thinking and GTC-2.5 Agent are already in development, pushing the series toward deeper reasoning and autonomous execution.