News

We are thrilled to announce the release of 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.

At 400 million parameters, GTC-2.5 represents a significant scale-up from our earlier 64M Preview models. Trained on a curated 19GB dataset combining high-quality Chinese and English text, code, and mathematical content, it delivers substantially improved coherence, knowledge retention, and instruction-following capabilities.

GTC-2.5 mini (64M parameters) is now available under a permissive license on Hugging Face and GitHub — perfect for research, experimentation, and lightweight edge applications.

This release marks the end of the "Preview" era for GTC and the beginning of a more mature, product-oriented phase.

📖 Read full article →

We are excited to introduce GTC-2.5o, the first truly multimodal member of the GTC family. Built on the GTC-2.5 foundation, GTC-2.5o extends with full multimodal capabilities — text, image, and audio understanding — in a unified model.

Integrating vision and audio encoders directly into the 64M-parameter backbone, GTC-2.5o supports image captioning, visual question answering, audio comprehension, and cross-modal reasoning. It validates TAI Research's multimodal pipeline and serves as a baseline for larger vision-language-audio models.

The model is currently in early preview, available to Pro users on Sigma. We are actively collecting feedback and plan to iterate rapidly.

📖 Read full article →

We are excited to announce the release of GTC-2.5V Preview, the first vision-language foundation model developed by TAI Research. Extending the GTC series with multimodal capabilities, this 64M-parameter model integrates a lightweight visual encoder with a decoder-only language backbone, trained end-to-end on a V100 32G GPU.

GTC-2.5V Preview demonstrates foundational capabilities in image captioning, visual question answering, and text reading from images. While it is a compact research preview, it serves as a baseline for our upcoming larger vision-language models and downstream applications, including multimodal agents and Computer Use scenarios.

This release reaffirms our commitment to transparent, efficient AI research across modalities. We extend our gratitude to the MiniMind-V team for their open-source contributions, enabling us to explore multimodal AI efficiently.

📄 System Card Available: Download GTC-2.5V Preview System Card (PDF)

We are excited to announce the release of GTC-2.5 Preview, the first conversational foundation model developed by TAI Research. Built upon the open-source MiniMind project and trained on a V100 GPU, this 64M-parameter model marks a critical milestone in validating our technical pipeline.

GTC-2.5 Preview demonstrates foundational capabilities in dialogue, knowledge retrieval, and code generation. While it is a compact research preview, it serves as a baseline for our upcoming larger models and applications, including domain-specific AI assistants and lightweight edge deployments.

This release reaffirms our commitment to transparent, efficient AI research. We extend our gratitude to the MiniMind team for their open-source contributions, enabling us to stand on the shoulders of giants.

📄 System Card Available: Download GTC-2.5 Preview System Card (PDF)

We're proud to announce GTC-Guard-0, the first model in the TAI Guardian series. This lightweight text safety classifier (under 10M parameters) is built for interpretability and CPU inference. It marks TAI Research's first step into AI alignment and content safety — building values from the first line of code. Available now on Hugging Face under MIT license.

🤗 Download on Hugging Face →

After a long rest, we are now back. TAI Sigma Chat is now called Sigma.

We are back after a long pause. Weights and repository published.

The previous Sigma Gobang Turbo is too slow, so we create a new one — Sigma Gobang 2.

An Early Supervised Fine-Tuned GTC-2 model for improved conversational performance.

We enhanced the algorithm of Sigma Gobang and built the Sigma Gobang Turbo.

You can now access the Sigma Gobang online!

Developers can now visit Hugging Face to download these two models. The two models have the performance of the approximate GTC-2 Large while reducing the parameter size.

Users can explore the chat freely without an account, or choose to register and log in to unlock persistent conversation history. This marks a step forward in making TAI Sigma Chat not only a research demo, but also a growing platform for interactive intelligence.

Weights and repository published. Public beta available on the demo site.

You can now select different models and experience them.

Start using our models instantly!