On April 17th, China Telecom announced the open-sourcing of its massive AI model with 12 billion parameters—TeleChat-12B, marking another significant milestone in the domestic large-model technology development.
Compared to the 7B parameter version released in January, TeleChat-12B has achieved approximately a 30% overall improvement in content understanding, performance, and application scenarios. Particularly noteworthy is the over 40% enhancement in capabilities for multi-turn conversation reasoning and security-related areas.
Additionally, there are reports indicating that China Telecom plans to launch ultra-large AI models with parameter scales reaching hundreds of billions within this year.
To ensure high efficiency, TeleChat-12B has increased its training data volume from 1.5TB in the 7B version to 3TB, significantly enhancing both data quality and model performance.
Moreover, TeleChat-12B leverages actual interaction data from telecom users to optimize reward models and reinforcement learning models, effectively enhancing the model's performance in question-answering scenarios.
Currently, the TeleChat model has been applied in various practical scenarios including text creation, programming, network fault analysis, and business operations analysis. For instance, in text creation, the model generates an average text length exceeding 1500 characters, with an adoption rate as high as 85.7%.
For enterprise and institutional clients, the TeleChat model supports diversified scenarios such as business operations analysis, government consultation, and response to public grievances, covering 95% of actual business needs, with an accuracy rate of 90% in multi-turn conversation understanding.
Beyond traditional industries, the TeleChat model is actively exploring applications in more cutting-edge fields, such as making progress in intelligent drone choreography and lightweight construction.
