Beijing Dianzi Zhishu Technology Co., Ltd. (hereinafter referred to as 'BDNI') has recently achieved a significant milestone by successfully integrating the full-scale DeepSeek-V3/R1 model with domestic chips from Hygon DCU, Huawei, BiRen Technology, and Moxi, leveraging its 'Pagoda·Model Adaptation Platform'. This accomplishment provides developers with diverse computing power options when utilizing the DeepSeek model, allowing users to overlook underlying hardware differences when employing domestic computing power. Consequently, this facilitates rapid development, deployment, and model invocation.

DeepSeek-V3/R1 is an advanced AI model developed by DeepSeek, boasting robust data processing and analysis capabilities. It finds widespread application in natural language processing, image recognition, speech recognition, and other domains. Leveraging innovative engineering techniques such as the DeepSeekMoE hybrid expert system, multi-head latent attention mechanism (MLA), and a self-developed training framework, DeepSeek-V3/R1 achieves performance comparable to OpenAI GPT-1 with R1 and OpenAI GPT-4 with V3, while requiring fewer GPU training resources and time. The model excels in high inference efficiency and low training costs.

BDNI's 'Pagoda·Model Adaptation Platform' maintains high compatibility with major mainstream development frameworks and has adapted 24 base large models to date, further lowering the usage threshold for enterprises and developers. It seamlessly bridges chip architecture differences, enabling swift optimized support and hardware adaptation for models. For instance, in handling inference tasks, BDNI's mixed inference technology route aligns well with the DeepSeek technology route, optimizing inference effects, accelerating inference speed, reducing costs, and resolving insufficient computing power issues. With the support of BDNI's 'Pagoda·Model Adaptation Platform', DeepSeek-V3/R1 operates efficiently and stably on mixed chips, supporting diverse application scenarios.

Moreover, to fully harness the performance potential of domestic chips and model adaptation capabilities, BDNI has introduced the 'Spark·Domestic Computing Power AI Native Adaptation Certification'. This initiative strengthens the adaptation and coordination between domestic models and domestic computing power, fostering better support for AI native application scenarios by domestic chips.

In the synergistic interplay between domestic chips and outstanding domestic large models like DeepSeek, we are witnessing the dawn of full-stack AI localization...