This time last year, a model release from DeepSeek sent global tech stocks into a tailspin, earning the label "black swan."
Now V4 is here. Market reaction: okay.
1.6-Trillion-Parameter Open-Source King, No One Is Impressed Anymore
On April 24, DeepSeek dropped the V4 series on Hugging Face. Two versions:
| Model | Total Parameters | Active Parameters | Context |
|---|---|---|---|
| V4 Pro | 1.6 trillion | 4.9 billion | 1 million tokens |
| V4 Flash | 28.4 billion | 1.3 billion | 1 million tokens |
V4 Pro leaves Kimi K2.6's 1.1 trillion and its own V3.2's 671 billion in the dust, securing the throne as the world's largest open-weight model.
Pricing is even more aggressive. V4 Flash costs $0.14 input and $0.28 output per million tokens; V4 Pro costs $0.145 input and $3.48 output. Both undercut comparable models from OpenAI, Google, and Anthropic.
Benchmarks? TechCrunch's report is blunt: on reasoning benchmarks, V4 "nearly closes the gap with leading models," while on knowledge tests it "lags slightly behind" GPT-5.4 and Gemini 3.1 Pro, trailing frontier models by about 3-6 months overall. In coding competitions, it trades blows with GPT-5.4.
But Hong Kong Stocks Didn't Budge This Time
On April 27, the first trading day after DeepSeek V4's launch, market sentiment was a completely different story from last year.
In a Reuters report, Omdia chief analyst Lian Jye Su was measured:
"This release follows a fairly predictable path, as advances in model architecture and efficiency have been widely explored across the industry and academia. Expectations of new entrants are now priced into valuations."
In plain English: the shock threshold has changed.
When R1 dropped last year, Wall Street's logic was "how could China be this fast" — and Nvidia lost hundreds of billions in market cap overnight. This year? Over the past six months, Kimi K2.6 took first place on SWE-Bench Pro, Qwen3.6-Max topped six benchmarks, and Zhipu GLM-4.6 adapted to domestic chips. Open-source output from Chinese teams is now the norm. Another V4 Pro gets a "meh" from the market.
Artificial Analysis was equally restrained: V4 Pro shows significant improvements, but "overall, it ranks among leading open-weight models, not clearly surpassing competitors."
What Really Matters Is Not Benchmarks, but Huawei Chips
Alfredo Montufar-Helu, managing director at Ankura China Advisors, offered a line more important than any benchmark:
"Last year was the 'wow factor' — that's already priced in. What matters now is whether China can continue advancing in AI development, possibly using its own chips."
V4 is adapted for Huawei's Ascend. That's the real geopolitical weight of this release — whether US export controls can hold back Chinese AI. DeepSeek answered with one product iteration: at least on the algorithm and engineering front, it can't. The question is how far hardware can keep up.
A Few Takeaways Worth Remembering
First, the marginal wow factor of open-source large models is diminishing. This isn't to say DeepSeek did poorly — it's that market expectations have caught up with output. To recreate last year's "wow" moment, something else is needed — like actually surpassing closed-source models on some frontier task, not just narrowing the gap.
Second, V4 Flash's 1.3 billion active parameters + 1 million context + rock-bottom price may be more interesting to API developers than Pro. This combination's cost-effectiveness in long-document processing and agent backend orchestration will make a lot of mid-tier closed-source APIs look bad.
Third, the question "Can China do AI with its own chips?" doesn't have a full answer in 2026, but V4 brings the answer closer to "yes."
As for the market — it's already priced this in. If you want to see a show again, wait for real-world V4 Pro benchmarks on long-chain tasks like agentic coding.
That's when a new story begins.
Sources: CocoLoop; DeepSeek previews new AI model that 'closes the gap' with frontier models (TechCrunch); DeepSeek's new AI model does not wow markets in fast-changing industry (BusinessWorld / Reuters)