Unverified reports circulating on 13 August suggest Chinese artificial-intelligence firm DeepSeek may introduce its second large language model, DeepSeek R2, between 15 and 30 August. Leaked technical notes indicate the system is trained on a Huawei Ascend 910B cluster delivering about 512 petaflops of FP16 compute with 82% chip utilisation, or roughly 91% of the efficiency of Nvidia’s A100 GPUs. Commentators also claim the mixture-of-experts model could operate at about 97% lower cost than OpenAI’s GPT-4. The company has not issued any statement, and the rumoured timetable was contested by a person described by financial news outlet Sino Market as being close to DeepSeek, who said there are no plans for an August release. The conflicting accounts leave the launch schedule and exact capabilities of DeepSeek R2 uncertain.