HongYun released HyperLM V2.5, a new version of its combined training and inference platform for large models, according to IT Home and ifeng Tech reports on November 3. The Beijing-based company says it is the first to complete deep adaptation to Huawei's Ascend 950DT, is fully compatible with the air-cooled Atlas 850E SuperPoD and has passed Huawei's Ascend Native technical certification. HyperLM V2.5 supports Huawei's UnifiedBus 2.0, which the company says releases 13.4TB/s of system interconnect bandwidth, and is optimized for mxFP4 inference to use the 14.2 PFLOPS of a single node. It supports the latest open-source Qwen, GLM and DeepSeek models, and its scheduler has been rebuilt around the 950DT's memory and bandwidth profile. The release adds five capabilities: prefill-decode disaggregation, multi-level KV caching, vNPU virtualization that slices Ascend 950DT, 910C and 910B chips with fine-grained quotas and isolation for AI Core compute and HBM memory plus hot migration and automatic rescheduling, a unified model gateway compatible with the OpenAI protocol, and full-stack monitoring. For regulated sectors such as government, energy, transport and finance, the platform supports separation of administrator, security and audit roles, tiered classification controls and full operation logs, along with compute metering, model call auditing and per-project usage