In a landmark shift for the global AI landscape, DeepSeek is reportedly building its next-generation V4 model to run entirely on Huawei Ascend AI chips, according to a report from The Information. This marks a significant departure from the industry standard of using NVIDIA hardware and signals that China’s domestic “parallel tech stack” is reaching frontier-level maturity.
The move follows months of quiet collaboration between DeepSeek, Huawei, and Chinese chipmaker Cambricon to rewrite core code components to bypass NVIDIA’s CUDA ecosystem in favor of Huawei’s CANN architecture.
1. Technical Blueprint: DeepSeek V4
Insiders suggest V4 is not just an incremental update but a structural overhaul designed for “domestic-first” efficiency.
| Feature | Specification (Projected) | Significance |
| Parameters | 1 Trillion (1T) | A 50% increase over V3’s 671 billion. |
| Architecture | MoE + Engram Memory | Uses “Engram” vectors for long-term persistent memory across sessions. |
| Context Window | 1 Million Tokens | Enabled by the new “mHC” (Multi-head Conditional) attention mechanism. |
| Hardware | Huawei Ascend 950PR | Optimized for Huawei’s latest FP8-capable AI silicon. |
| Inference Speed | 1.8x Faster | Claims significant latency gains over V3 despite the larger size. |
2. The “NVIDIA-Free” Strategy
While Chinese giants like Alibaba and Tencent still maintain massive NVIDIA H20 clusters, DeepSeek is the first major lab to “divorce” from U.S. chipmakers for a flagship release.
- Code Rewrite: DeepSeek reportedly spent the first quarter of 2026 working with Huawei engineers to port their Multi-head Latent Attention (MLA) and DeepSeekMoE frameworks to run natively on Ascend NPUs.
- The Performance Gap: While individual Huawei chips (like the Ascend 910C) still trail NVIDIA’s H100 in raw peak performance, DeepSeek is compensating through architectural efficiency, proving that “smart” software can overcome “slower” hardware.
3. Market Impact: The “Two-Track” Era
The report has reignited the debate over the U.S. export ban’s effectiveness.
- Domestic Sufficiency: Analysts at IDC note that Chinese chipmakers captured 41% of the local AI accelerator market in 2025. DeepSeek V4 is seen as the “validation” that China can now train near-frontier models without TSMC-fabricated NVIDIA silicon.
- Stock Market Ripple: When the news broke, it triggered a brief sell-off in U.S. semiconductor stocks as investors questioned the long-term “moat” of Western AI hardware in the Asian market.
4. Release Timeline: “The Coming Weeks”
DeepSeek is currently in the final “stress-test” phase for V4.
- Primary Launch: Expected in mid-to-late April 2026.
- Variants: The company is reportedly working on two additional V4 variants specifically optimized for localized inference on consumer-grade Chinese GPUs (like those from Moore Threads and Biren).
5. Why It Matters: Coding & Reasoning Supremacy
Early internal benchmarks leaked alongside The Information report suggest that V4 aims to surpass Claude 4 and GPT-5.2 in specific technical domains:
- Automated Refactoring: The ability to rewrite entire legacy codebases into modern languages.
- Scalable Error Detection: Identifying logical flaws in million-line repositories.
- Architecture Planning: Generating full system designs from natural language prompts.
“DeepSeek V4 is the first visible sign that frontier model development on Huawei Ascend is already a reality,” noted tech analyst Max Song. “The world is about to see what happens when the most efficient AI lab in the world pairs with the most ambitious domestic hardware program.”