DeepSeek V4 is Coming: 1T Parameters, Huawei Ascend, Chinese LLMs Stand Tall

This is pretty interesting.

Reuters reported on April 10 that DeepSeek is expected to release its latest large model V4 in late April, featuring around 1 trillion parameters, using Mixture of Experts (MoE) architecture, and confirmed to run on Huawei Ascend processors.

Why do I say “interesting”? Because this is the first Chinese AI model reaching frontier performance that will be completely independent of NVIDIA GPUs.

What Does 1 Trillion Parameters Mean?

Don’t rush to say “what’s the use of large parameters.” I know many people are tired of the parameter race, but DeepSeek V4’s 1 trillion parameters are in the same ballpark as GPT-4.

More importantly, DeepSeek uses MoE (Mixture of Experts) architecture. The core idea: not all parameters are activated during every inference. Instead, a subset of “expert” networks is dynamically selected based on input. In other words, while the model has 1 trillion parameters, actual inference might only use tens of billions—preserving large model capabilities while controlling compute costs.

This is why DeepSeek can price so aggressively: big model, but low inference cost.

Huawei Ascend: From “Usable” to “Good”

What I’m more interested in is the “Huawei Ascend processor” detail.

Over the past few years, Chinese large model training and inference basically relied on NVIDIA GPUs. It’s not that people didn’t want to use domestic chips, but there were gaps in software ecosystem, compute stability, and ease of use. Training large models on Huawei Ascend used to be more “politically correct” than “technically optimal.”

But DeepSeek V4 confirmed to run on Huawei Ascend shows that Huawei’s chip capabilities have evolved from “usable” to “good”—at least for frontier large models like DeepSeek, Ascend can meet requirements.

This is no less significant than the model itself. Because it means: Chinese large models finally have their own “compute foundation.”

DeepSeek’s “Value” Strategy

Honestly, I’ve been following DeepSeek for a while.

Their strategy is clear: offer near-frontier performance at far lower prices than competitors. DeepSeek V3 already proved this—performance close to GPT-4, but at one-tenth the price.

If V4 really achieves frontier performance while running on Huawei Ascend, the cost advantage will be even more obvious. Because Huawei chips are definitely cheaper than NVIDIA GPUs.

My personal judgment: DeepSeek is using a “value” strategy to force the entire large model industry to lower prices. Good for users, but huge pressure for competitors.

Chinese LLM “Independence”

Here I’ll make a potentially controversial point: DeepSeek V4 running on Huawei Ascend is more important than its performance.

Why? Because while Chinese large models have performed well, their training and inference relied on NVIDIA GPUs. If US sanctions escalate and cut off high-end GPU supply, Chinese large models would face “compute stranglehold” risk.

But DeepSeek V4 running entirely on Huawei Ascend means Chinese large models finally have “independence”—not constrained by external supply chains, able to control their own compute foundation.

This value might not show in the short term, but long-term, it could reshape the entire large model industry landscape.

Final Thoughts

DeepSeek V4 hasn’t launched yet, so saying “Chinese LLMs finally stand tall” might be premature. But at least, I see a signal: Chinese large models are shifting from “chasing quantity” to “chasing independence.”

This isn’t just a technical issue—it’s a strategic one.

Next, I’ll keep watching V4’s actual performance after launch. If it really reaches frontier level, then Chinese large models will truly have found their footing.