Chinese artificial-intelligence developer DeepSeek is preparing to release an upgraded large language model, DeepSeek R2, in the second half of August, according to several reports circulating in industry media. The company has not issued an official statement, but multiple accounts place the launch window between 15 and 30 August. The new model is said to be trained on a cluster of Huawei Ascend 910B processors, delivering about 512 petaFLOPS of FP16 performance and operating at roughly 91 per cent of the efficiency of Nvidia’s A100 GPU. Sources familiar with DeepSeek’s plans add that the system utilises around 82 per cent of available processor capacity, a figure that would imply improved hardware utilisation compared with earlier iterations. While technical specifications remain unverified, early marketing material suggests DeepSeek R2 could run inference workloads at costs up to 97 per cent lower than OpenAI’s GPT-4. The model is expected to follow a mixture-of-experts architecture and to be released under an open-source licence, mirroring DeepSeek R1’s distribution strategy.