DeepSeek V4 Incoming: Trillion-Parameter Open-Source AI Model to Run on Huawei Chips

Chinese AI company DeepSeek is preparing to release its V4 model later in April. The trillion-parameter Mixture-of-Experts model will be released as open source under the Apache 2.0 license, marking a major milestone in open AI development.

Technical Specifications

DeepSeek V4 scales to approximately 1 trillion total parameters but activates only 37 billion per token, keeping inference costs comparable to V3. A 1-million-token context window is powered by the Engram conditional memory architecture, achieving 97% accuracy on the Needle-in-a-Haystack benchmark at million-token scale.

Native Multimodality

Unlike models that bolt on vision capabilities, V4 integrates text, image, and video generation during pre-training, enabling more coherent cross-modal reasoning.

Huawei Hardware

Reuters confirmed on April 4 that DeepSeek V4 will run on Huawei's Ascend 950PR chips. This represents the first frontier AI model built to run on Chinese semiconductor infrastructure, a significant development amid ongoing US-China tech competition.

Competitive Landscape

DeepSeek V4 will compete with OpenAI's GPT-5.4, Anthropic's Claude Mythos 5, and Google's Gemini models. Its open-source nature is a key differentiator. The model reportedly scores 81% on SWE-bench, signaling strong performance in software engineering tasks.