DeepSeek V4 Countdown: What Surprises Can Chinese Models Still Deliver?

April’s AI news cycle has been dominated by big tech announcements. But there’s another “sleeper” story: DeepSeek V4 is coming.

Sources say founder Liang Wenfeng confirmed internally that V4 launches in late April. The timing—just before May Day holiday—catches the tail end of April’s release wave while leaving room for holiday buzz.

Honestly? I’ve always been fascinated by DeepSeek.

Not because their tech leads—it does, V3 was impressive—but because of their “atypical” path. While AI labs race on parameters, compute, and funding, what did DeepSeek focus on? Efficiency.

V3 cost $5.57 million to train. In Silicon Valley, that might cover OpenAI’s coffee budget for a week. Yet DeepSeek built a model matching GPT-4o on multiple benchmarks with that budget. This “do more with less” capability? That’s what’s truly scary.

Now V4 is coming. The big question: what breakthroughs does it bring?

Leaked info points to three focus areas:

First, reasoning. Not just “answer questions,” but “complex problem decomposition + multi-step reasoning.” Basically, making AI think like humans, not just retrieve like databases.

Second, code. V3 was already strong, but V4 is supposedly “specially optimized.” What does this signal? DeepSeek is targeting developers as core users. Smart strategy—developers are pickiest yet most influential.

Third, multimodal. Text, images, code—seamless switching. Not a new concept, but hard to execute. If V4 delivers here, it’s not just “China’s pride”—it’s “world-class.”

But what I anticipate most isn’t the tech—it’s whether DeepSeek continues its “price butcher” approach.

V3’s API pricing: mere cents per million tokens, flooring the entire industry. What will V4 do? Keep subsidizing for market share, or start monetizing? The answer may signal Chinese AI’s commercialization path.

Another suspense: will V4 be open source?

V3 was open source—that’s how it built developer community buzz fast. Will V4 continue? Open source means transparency, but also lets competitors cheaply replicate. For DeepSeek, a startup, this is a tough call.

My guess: V4 probably open sources, but with a “delayed window”—API service first to grab market, model release months later. Maintains developer anticipation while securing revenue.

One interesting detail: DeepSeek’s Hangzhou office sits near Alibaba. Yet every time big tech launches, DeepSeek plays “late mover”—not first to announce, but first to win mindshare through quality. This “half-step slower but more solid” style apparently matches Liang Wenfeng’s personality.

Will V4 make us go “wow” again? I think so. Not because DeepSeek is that dominant, but because Chinese AI ecosystems have reached a tipping point where quantity transforms to quality.

What capability should V4 prioritize? Reasoning, code, or multimodal?