DeepSeek V4 Series Models Open-Sourced with 1M Context Length, Huawei Ascend Provides High-Performance Support

DeepSeek V4-Pro and V4-Flash have been officially open-sourced, with context length expanded from 128K to 1M. New KV Cache sliding window and compression algorithms are added, significantly reducing computation overhead and better supporting Agent and Coding scenarios.

Huawei Ascend series products fully support the models: Ascend 950 super node achieves low-latency inference (V4-Pro 20ms, V4-Flash10ms), and A3 super node provides high-throughput deployment; Huawei Cloud MaaS platform supports one-click API calls.

In addition, Ascend CANN launched the PyPTO programming paradigm and TileLang scheme open-source, lowering the threshold for operator development and shortening the model landing cycle.