DeepSeek-V4, the Chinese AI model adapted for Huawei chips

• V4 comes in two versions: Pro (higher performance, up to 12 times more expensive due to compute constraints) and Flash (faster, cheaper, with weaker capabilities on demanding tasks)
• Huawei confirmed its Ascend 950 chips were used for part of V4-Flash's training, marking DeepSeek's pivot from Nvidia chips used in earlier V3 and R1 models
• DeepSeek said Pro pricing could fall sharply once Huawei Ascend 950 supernodes are deployed at scale in the second half of 2025, though analysts note Huawei still trails Nvidia technologically

Summary

Chinese AI startup DeepSeek released a preview of its V4 model on April 24, marking a significant advancement in China's AI self-sufficiency efforts. The model is specifically adapted to run on Huawei's Ascend AI chips, distinguishing it from earlier versions (V3 and R1) that were trained on Nvidia hardware.

Key Model Features:

• Two versions available: Pro (high-performance, expensive) and Flash (faster, cheaper, lighter)
• Both support 1-million-token context window
• Designed for agent frameworks including Claude Code and OpenClaw for complex, multi-step tasks
• Pro excels in agentic coding, STEM, and competitive programming, outperforming all open-source models in maximum reasoning mode
• Still trails Google's Gemini 3.1 Pro and OpenAI's GPT-5.4 in some areas

Strategic Partnership:

Huawei confirmed V4 runs on Ascend 950-based supernode clusters, with chips used for partial V4-Flash training. The collaboration demonstrates performance parity between Huawei and Nvidia hardware, lowering barriers for Chinese developers to build entirely domestic AI solutions.

Market Implications:

The partnership represents tangible progress toward China's AI infrastructure independence amid U.S. export controls on Nvidia chips. However, analysts note Huawei still trails Nvidia technologically. Pro version currently costs up to 12 times more than Flash due to compute capacity constraints, though DeepSeek expects pricing to drop significantly when Ascend 950 supernodes deploy at scale in H2 2025.

Bottom Line:

DeepSeek-V4 demonstrates China's advancing capability to develop competitive AI models on domestic hardware, though challenges remain in matching frontier Western systems and achieving full technological independence.