MWS AI

Rubytech Tests Russian LLM Cotype 3 on Chinese GPUs Matching NVIDIA H100 Performance

Rubytech has successfully tested the Russian large language model Cotype 3 developed by MWS AI on Chinese graphics accelerators instead of NVIDIA hardware. The trials took place on the Skala^r AI machine software-hardware complex and demonstrated stable operation with inference performance comparable to configurations using NVIDIA H100 GPUs. Eight Chinese GPUs were used in the test system, delivering an average first-token generation time of about 8 seconds and inter-token latency between 111 ms when processing 27,000-token contexts. Engineers adapted drivers, software environments, and orchestration tools, achieving up to 2-2.2 times faster performance in certain scenarios compared to the standard Chinese GPU setup. The results indicate that corporate AI infrastructure can be built without mandatory reliance on NVIDIA accelerators, which is particularly relevant for government agencies, state-owned companies, and critical information infrastructure organizations requiring localized model deployment and controlled scaling. Rubytech also noted potential reductions in total cost of ownership with Chinese GPUs, although the company emphasized this is an additional option rather than a complete replacement for NVIDIA solutions.

Russia Simplifies State Support Access for National AI Model Developers

Russian authorities have decided to shorten and clarify the path to government support for developers of large AI models. Following the entry into force of the law on artificial intelligence development, obtaining the status of a national or sovereign model will become easier, with decisions verified through a single set of test tasks. The reference test is planned to be published in open access and updated regularly, allowing developers to know in advance the exact criteria the state will use to evaluate their neural networks. Companies such as MWS AI and T-Bank will be able to apply for the new statuses and associated support measures. Expertise will be entrusted to several organizations that have passed state verification, with the main criterion being Russian company control over the entire model lifecycle rather than the origin of every line of code. The use of foreign components under open licenses will be permitted if the developer can independently modify, develop, and maintain the solution. Bureaucratic procedures will be reduced, missing documents can be submitted after the application, and computing infrastructure must be located in Russia but can be rented. The first areas of mandatory application of domestic models will be education and public services, with key provisions of the law taking effect on September 1, 2026, and requirements for sovereign models on March 1, 2027.

DeepSeek User Conversations Appear in Google Search Results via Publicly Shared Links

Conversations between users and the Chinese AI service DeepSeek, including Russian-language exchanges, have surfaced in Google search results. The exposed pages belong to Shared Conversations that users themselves made public through shareable links. These pages display full question-and-answer histories along with the names of any uploaded documents. No actual breach of DeepSeek occurred, and security researchers note that search engines indexed similar public chatbot dialogues more than a year ago. MWS AI confirmed that no closed chat histories or account access were exposed. Meanwhile, Yandex stated that links to conversations with its Alice AI assistant are blocked from indexing, remain active for only 14 days, and do not transmit attached files to recipients.