DeepSeek officially released its fourth-generation model, DeepSeek-V4, on Friday, April 24, 2026. This release is a major strategic pivot for the Chinese AI lab, as the model was explicitly adapted to run on Huawei’s Ascend AI chips, moving away from the company’s previous reliance on Nvidia hardware.
The launch is being hailed as a “Sputnik moment” for China’s domestic AI ecosystem, proving that frontier-level performance can be achieved using entirely homegrown hardware.
1. DeepSeek-V4 Product Lineup
DeepSeek released the model in two distinct versions, both available as open-source “preview” versions under an MIT license.
| Model | Parameters (Total) | Parameters (Active) | Primary Focus |
| DeepSeek-V4-Pro | 1.6 Trillion | 49 Billion | High-end reasoning, coding, and STEM tasks. |
| DeepSeek-V4-Flash | 284 Billion | 13 Billion | Cost-efficiency and low-latency inference. |
- Context Window: Both models support a massive 1-million-token context window, putting them on par with Google’s Gemini-3.1-Pro.
- Performance: In coding benchmarks, V4-Pro scored an 81% on SWE-bench Verified, trailing only the latest models from OpenAI and Google.
2. The Huawei “Ascend” Integration
Hours after the model’s release, Huawei confirmed that its entire Ascend SuperPod product line now provides full support for DeepSeek-V4.
- Hardware Target: The models are specifically optimized for the Ascend 910C and the newer Ascend 950PR chips.
- Engineering Feat: DeepSeek reportedly spent months rewriting core code to transition from Nvidia’s CUDA architecture to Huawei’s CANN toolkit.
- Training Role: Huawei stated that its chips were used for a portion of the training process for V4-Flash, marking a significant shift in how Chinese frontier models are developed.
- Cost Efficiency: Running the model on Ascend clusters reportedly reduced API costs by up to 97% compared to previous pipelines using Western models like GPT-4o.
3. Market and Geopolitical Impact
The collaboration is a direct response to tightening U.S. export controls on advanced AI chips.
- Stock Surge: Following the announcement, domestic Chinese chipmakers saw a rally: Hua Hong Semiconductor rose 15%, while SMIC advanced nearly 10%.
- Competitive Pressure: The launch caused “competitive anxiety” for other Chinese AI firms, with shares of rivals like MiniMax and Zhipu AI dropping 8-9%.
- The “Nvidia Risk”: Nvidia CEO Jensen Huang recently warned that a scenario where “DeepSeek comes out on Huawei first” would be a major challenge for the U.S. competitive lead in AI.
4. Architectural Innovations
DeepSeek-V4 introduced several “breakthrough” techniques to handle its massive 1.6-trillion parameter size on domestic hardware:
- Hybrid Attention: Combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), reducing KV cache requirements by 90% compared to previous versions.
- mHC (Manifold-Constrained Hyper-Connections): A new mathematical framework that prevents training divergence in massive Mixture-of-Experts (MoE) models, ensuring stability at the 1.6T parameter scale.
