DeepSeek & the start of the Chinese era of AI

DeepSeek, China, & the future of AI

As promised a few days ago, this post is going to focus on the nothing-short-of-crucial release of DeepSeek v4 from the homonymous company, because it reveals a lot about the state of Chinese AI and, overall, the future of the industry.

How much behind are they? Are they behind at all? What’s the long-term plan?

We’ll answer all these questions today by presenting the main insights from the paper. If you read this piece in full, you’re going to:

Fully understand how frontier AI models work and how they are built, with no word more complex than the previous,

Capturing subtleties about AI software and hardware that almost no one understands,

And comprehend the nuanced geopolitical implications of this model. There’s a reason the entire industry holds its breath whenever DeepSeek drops a new model, especially considering the last time they had a big model release, NVIDIA crashed 17%, half a trillion dollars, the largest fall on record, based on a paper by these guys.

I mean, what other non-US model could cause the level of panic this AI Lab consistently creates, even inside the US Government?

This piece will leave you with a bittersweet taste, especially if you’re concerned about US supremacy. Not because China has “caught up”, we are way past that simplistic conversation that judges who’s ahead based on who scores higher in contaminated benchmarks.

The real battle is being fought elsewhere, and I can’t help but feel quite worried about the US's chances if things don’t change before it’s too late.

You’re about to learn a ton about frontier AI today, so let’s dive in.

Redefining the open frontier

First, let’s start with the official release numbers, which, in themselves, are very impressive for an open-source model.

DeepSeek has released two new extremely cheap open models: DeepSeek-V4-Pro and DeepSeek-V4-Flash.

V4-Pro is the flagship model, with 1.6 trillion total parameters and 49 billion activated per token. That means only 3% of the model is active for each prediction (i.e., a huge model running like a “small” one).

V4-Flash is the cheaper, smaller variant, with 284 billion total parameters and 13 billion activated per token. Much less sparse relative to the other, but much smaller.

Both support context windows of up to 1 million tokens (i.e., they can handle ~750k words in context, so you can feed them a lot of information), making the release especially well-suited for long-context reasoning, coding, search, and agentic workflows.

But let’s dissect its claims more. On general knowledge and factuality benchmarks, DeepSeek-V4-Pro-Max, the highest-reasoning version of V4-Pro, shows a major jump over previous open models but still trails the strongest proprietary systems in some categories.

It scores 57.9 on SimpleQA Verified, compared with 46.2 for Claude Opus 4.6, 45.3 for GPT-5.4, and 75.6 for Gemini-3.1-Pro.

On Chinese-SimpleQA, it reaches 84.4, close to Gemini-3.1-Pro’s 85.9 and ahead of Claude Opus 4.6 and GPT-5.4. On MMLU-Pro, a broad academic knowledge benchmark, it scores 87.5, behind Gemini-3.1-Pro’s 91.0…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.