SelterRussia’s largest independent IT infrastructure provider updated AI model directory (Base Model Catalog) for handling large language models. The updates enhance the platform’s capabilities for enterprise customers: the company provides private access to models, expands the catalog with advanced generation models, and expands the ability to build geographically distributed systems with the addition of new RU-6 regions. These changes will enable enterprises to roll out artificial intelligence solutions faster, reduce operating costs and comply with information security standards when handling sensitive data.
The AI model directory is contained in Selectel Artificial Intelligence Platform – An ecosystem of services for training and running generative models to help companies accelerate the integration of artificial intelligence solutions. This enables enterprises to use large language models, agent-based development tools and other artificial intelligence technologies without having to build their own IT infrastructure, greatly reducing technical barriers to innovation.
Image source: Selectel
The service is now available in new fault tolerance zones based on three separate Availability Zones. Customers can deploy two identical inference services in different availability zones to establish a multi-region solution and reduce the risk of downtime due to geographical distribution of infrastructure.
When building an inference service, you can now choose the type of access: public (over the Internet) or private (access only within the user’s private network, no traffic to the Internet). Companies will be able to use the Direct Connect service to establish a dedicated channel between their infrastructure and the Selectel data center and send requests to the AI model catalog through it. This will enable enterprises with higher security requirements to use models in isolated loops, which is especially important for financial, healthcare and government organizations that handle sensitive data.
Table of contents Supplemented with a new model in NVFP4 quantization which improves performance but slightly reduces accuracy. This technology is optimized for Blackwell generation GPUs, allowing you to run resource-intensive models on compact configurations. Queen series models have a maximum context size of 262,000 tokens, which allows you to handle large files and long agent scenarios while maintaining high accuracy. In the near future, other popular models will appear in FP8 and NVFP4 quantization, optimized for the GPU, and these models will be available in the new RU-6 region.
Advertisement | AO “Selektel” INN 7810962785 erid: F7NfYUJCUneVeSH3AA3q
source: