DeepSeek V4 Is Coming: Trillion Parameters on Huawei Ascend—What Does It Mean?

Word has been quietly circulating in AI circles: DeepSeek V4 is expected to drop in late April.

One trillion parameters, MoE architecture, and most importantly—confirmed to run on Huawei Ascend processors.

The more I think about this, the more interesting it gets. Not because of the parameter count, but because it means domestic AI chips are finally facing a “real combat” test.

First, let’s look at DeepSeek V4’s fundamentals

What does 1 trillion parameters mean? OpenAI never officially disclosed GPT-4’s parameters, but industry consensus puts it around 1.8 trillion. If DeepSeek V4 really hits 1 trillion, it’s already knocking on the door of the first tier.

MoE (Mixture of Experts) architecture is now standard. Basically, it splits a large model into many small experts, activating only a subset each time—saving compute while maintaining performance. DeepSeek was already using MoE in V3, so V4 is likely an evolution of that approach.

But what I care about most isn’t these technical specs—it’s those four words: “Huawei Ascend.”

Can the domestic chip + frontier model combination actually work?

Honestly, the relationship between domestic AI chips and frontier large models has been awkward. On one hand, major model vendors all train on NVIDIA GPUs. On the other, while domestic chips “work,” they’ve mostly been limited to inference deployment or running “watered-down” model versions.

If DeepSeek V4 is really training and inferencing on Ascend, that’s a milestone—the first time domestic chips are running a trillion-parameter frontier model.

I asked several friends working in AI infrastructure, and they generally see this as having two sides.

The good: it proves Ascend’s compute density and interconnect bandwidth are actually sufficient. Trillion-parameter models have extreme hardware requirements, especially for multi-card communication. If DeepSeek can run on Ascend clusters, it shows domestic chips have made a qualitative leap in engineering capabilities.

The concern: the optimization workload could be massive. CUDA has accumulated years of maturity—all sorts of operator libraries, communication libraries. Ascend’s CANN ecosystem has improved quickly, but completely painless migration? Probably not realistic. The DeepSeek team likely spent enormous effort on low-level adaptation and operator optimization.

What does this mean for the domestic AI ecosystem?

I think it could be a turning point. Previously, everyone saw domestic chips as “backup options”—only remembered when NVIDIA played hardball. But if DeepSeek V4 on Ascend actually delivers, that’s different—developers will seriously consider using Ascend as primary, not just a temporary fix.

Plus, DeepSeek has a reputation for open-sourcing. V3 was open-source; V4 likely will be too. If the open-source code includes complete Ascend adaptation solutions, that’s a huge contribution to the entire community. Other teams can reference it directly instead of stumbling through the same pitfalls from scratch.

Of course, there’s a concern: performance comparison.

If DeepSeek V4 on Ascend achieves what percentage of the NVIDIA solution’s inference speed and training efficiency? 80%? 90%? Or only 60%? This number will directly impact market choice. In commercial scenarios, compute is cost—a 10% efficiency gap could mean life or death.

From a broader perspective

Against the backdrop of US-China AI competition, chip and model binding is getting deeper. With US restrictions on high-end GPU exports, China needs its own alternatives. The DeepSeek V4 + Huawei Ascend combination can be seen as an important breakthrough attempt by China’s AI industry under “decoupling” pressure.

Whether the breakthrough succeeds is still uncertain. But at least someone’s seriously doing it, not just talking.

I’m quite looking forward to the benchmark comparisons after V4 launches. Not because I’m pessimistic—I just think “running” and “running well” are two different things. I hope DeepSeek can produce convincing data.

One final thought: regardless of the outcome, this attempt itself deserves respect. It’s certainly better than certain people who just say “domestic stuff doesn’t work” and then lie down waiting to die.