The Darwin-397B-ZTC model with self-checking technology is presented

Developers presented Darwin-397B-ZTC — a new MoE model with 397 billion parameters, built on the Qwen 3.5 architecture, the system uses FP8 precision and is focused on solving complex logical tasks. The key feature of the model was the Zero-Token Confidence (ZTC) technology, this mechanism allows the neural network to conduct internal self-checking and evaluate confidence in the response before the first token of the text is generated, such an approach is aimed at reducing the likelihood of hallucinations when working with critical data. During testing, the model showed a result of 93.43% in the GPQA Diamond benchmark, which evaluates AI capabilities in the field of expert scientific knowledge, the system's performance was 19 points when performing tasks in 2 hours, which confirms the effectiveness of the architecture in processing high-load requests. The model is available for review and integration through the repository on the Hugging Face platform, the use of FP8 allows optimizing video memory consumption while maintaining the accuracy of calculations, which makes the system suitable for deployment in specialized computing environments. The introduction of ZTC changes the approach to generative inference, moving from simple probabilistic generation to an iterative verification process, this allows using the model in scenarios where the accuracy of the response is more important than the speed of text generation.