DeepSeek V4 Anticipated for Late April Launch with 1 Trillion Parameters
China's DeepSeek V4 runs on Huawei Ascend chips, bypassing US export controls with 1M token context.
DeepSeek is preparing to launch its V4 model in late April 2026, according to reports from Reuters and The Information. The model reportedly features approximately 1 trillion total parameters using a Mixture-of-Experts (MoE) architecture, with only 32-37 billion parameters active per token—keeping inference costs manageable despite the massive scale. While not officially confirmed, leaks suggest V4 could support a 1 million token context window, building on DeepSeek's Engram conditional memory system research from January 2026. The model has already been delayed twice, but recent infrastructure testing with the smaller V4-Lite variant suggests the full release is imminent.
Most significantly, Reuters confirmed V4 will run exclusively on Huawei's Ascend processors, making it the first frontier-class AI model built around Chinese domestic semiconductor infrastructure. DeepSeek deliberately denied NVIDIA and AMD early optimization access while giving Chinese chipmakers priority—a strategic move signaling China's push for AI hardware independence amid US export controls. If benchmarks hold up, this demonstrates that China's AI progress isn't meaningfully slowed by restrictions on Western hardware. The combination of frontier performance, aggressive DeepSeek pricing, and domestic chip optimization could send shockwaves through the global AI market.
- 1 trillion parameter Mixture-of-Experts model with only 32-37B active parameters per token
- Runs exclusively on Huawei Ascend chips, bypassing US export controls on NVIDIA/AMD hardware
- Potential 1 million token context window building on Engram memory system research
Why It Matters
Demonstrates China's AI independence from Western chips, potentially reshaping global semiconductor competition and AI accessibility.