DeepSeek V4 Preview Drops, and This Time It’s Running on Chinese Chips

DeepSeek just dropped a preview of V4, the follow-up to the model that sent shockwaves through Silicon Valley a year ago. And this time, they’re making a pretty bold claim: it can hang with the big dogs—GPT-5, Claude 4, Gemini 2.5—without being locked behind a paywall.

I’ve been watching DeepSeek since the V3 days, and honestly, their pace is impressive. V4 isn’t just another incremental update. According to their release, the biggest leap is in coding. That’s the battleground right now—agents like ChatGPT Codex and Claude Code are making developers’ lives easier, and DeepSeek wants a piece of that action. If V4 can actually match those tools, it’s a serious contender.

But here’s the part that caught my eye: DeepSeek specifically called out compatibility with Huawei’s chips. That’s a big deal. For months, the narrative has been that China’s AI ambitions are hamstrung by US export controls on NVIDIA’s best GPUs. If DeepSeek can train and run a competitive model on domestic hardware, it’s not just a technical win—it’s a geopolitical flex. The Huawei Ascend lineup has been quietly improving, and this could be the proof point that changes how people think about China’s chip independence.

Now, this is just a preview. We don’t have full benchmarks yet, and I’m always skeptical of benchmark claims from companies that have something to prove. But the fact that they’re releasing it as an open-source model is a power move. OpenAI and Anthropic keep their weights locked down. DeepSeek is saying, “Here, try it yourself.” That builds trust, even if the model isn’t perfect.

I also appreciate that they’re not trying to hide the domestic angle. Some Chinese tech companies dance around it. DeepSeek straight-up said V4 is optimized for Huawei silicon. That’s refreshingly direct, and it signals they’re betting on a future where they don’t need TSMC or NVIDIA to compete.

Will V4 actually dethrone Claude 4? I doubt it, at least not out of the gate. But it doesn’t have to. If it’s 90% as good and free, that’s enough to disrupt the market. The real test will be when developers start running it on their own machines. I’ll be downloading the weights the moment they’re available.

One thing’s for sure: DeepSeek isn’t a one-hit wonder. They’re building a real R&D pipeline, and the rest of the industry should be paying attention.