TL;DR: China's domestic Ascend 910C AI cluster has completed full-parameter post-training for the 1.6-trillion-parameter DeepSeek-V4-Pro model, proving the viability of domestic hardware for frontier-scale AI development.
Summary: A domestic Chinese AI computing cluster utilizing Ascend 910C chips has finished the full-parameter post-training of DeepSeek-V4-Pro. The mixture-of-experts model features 1.6 trillion total parameters, with 49 billion active parameters. This achievement showcases the growing capacity of domestic hardware stacks to run post-training pipelines for extremely large-scale models.
Why it matters: This signals that hardware constraints may not prevent the continuous scale-up of open-source models trained on alternative hardware architectures. Builders should monitor the performance of models trained on these clusters as they could further drive down API and hosting costs.
Source: @Lonely__MH