DeepSeek V4 and the World Model Race: What Actually Matters

DeepSeek dropped a preview of V4 on Friday, and it’s worth paying attention to for more than the usual hype cycle reasons. This is their first flagship model optimized for Huawei’s Ascend chips, which is a pretty direct test of how far China can move away from Nvidia. The model is still open source, matches closed-source rivals from Anthropic, OpenAI, and Google on benchmarks, and handles much longer prompts thanks to a more efficient architecture.

That last bit is the sleeper feature. Longer context windows aren’t just a spec sheet flex—they matter for real-world applications like legal document analysis, codebase understanding, and any task where you need the model to remember what you said 50 pages ago. V4’s design makes this less computationally expensive, which is the kind of engineering improvement that actually changes how people use the thing.

But the open-source angle is where it gets interesting. DeepSeek has been positioning itself as the open-weight counterweight to the closed-source giants, and V4 keeps that going while also showing that you can achieve competitive performance without the latest Nvidia hardware. That’s a big deal for any country or company worried about supply chain dependencies.

Meanwhile, the world model conversation is heating up again. The idea is simple: large language models are great at manipulating symbols and text, but they’re terrible at understanding physics, spatial reasoning, and cause-and-effect in the real world. You can’t ask an LLM to fold laundry or navigate a crowded street because it has no internal model of how objects interact or what happens when you bump into a table.

Fei-Fei Li and Yann LeCun have been pushing world models as the path forward for robotics and embodied AI. The thinking is that if you can build a model that learns a compressed representation of how the world works—gravity, friction, object permanence, the fact that a cup falls if you push it off a table—you can then use that model to plan actions in the physical world. It’s a fundamentally different approach from the “just throw more data at it” school of thought.

The timing is interesting because we’re also seeing a lot of money flowing into AI infrastructure. Google is reportedly investing up to $40 billion in Anthropic, valuing the company at $350 billion. That’s not just a bet on Claude—it’s a bet on compute capacity. Both Anthropic and OpenAI are fighting for GPU clusters like they’re the last lifeboats on the Titanic. The arms race for hardware is real, and it’s driving valuations that would have seemed absurd three years ago.

On the regulatory front, China blocked Meta’s $2 billion acquisition of AI startup Manus, citing national security. Beijing called it a “conspiratorial” attempt to hollow out its tech base. This is part of a broader pattern where China is tightening its grip on AI firms that try to leave, and the US-China tech rivalry is escalating in ways that don’t benefit anyone. There’s no winner in this kind of competition—just a lot of wasted potential and duplicated effort.

And in a move that’s getting less attention than it should, President Trump fired the entire National Science Board. The NSF has been a quiet but crucial player in developing foundational technology, from the internet to modern AI. Firing the whole board raises legitimate concerns about political interference in US science. It’s the kind of story that feels like a one-day headline but has long-term consequences for research funding and direction.

World models are on MIT Technology Review’s list of 10 Things That Matter in AI Right Now, and for good reason. If we want AI to actually help with physical tasks—manufacturing, elder care, disaster response—we need models that understand the world, not just the internet. DeepSeek V4 is a solid step forward for language models, but the real frontier is still ahead of us.