The Chinese artificial intelligence company DeepSeek has rolled out preview versions of its V4 series of models, marking the lab's most ambitious release since its V3 launch caused a sharp sell-off in US tech stocks early last year. The new V4 Flash and V4 Pro models are being positioned as the most powerful open-source AI platform yet built, and the launch comes at a moment when the strategic balance between Chinese and American AI labs is closer than at any point since the boom began.
Background and Context
For most of the modern AI era, the leading models have come from a small group of American labs, with OpenAI, Anthropic, and Google each holding the top spot at various points. That picture began to shift in early 2025, when DeepSeek's R1 model briefly matched the strongest US system on widely watched benchmarks. The performance was striking on its own terms, but the company's approach to releasing its model weights and training details turned the moment into a strategic event rather than just a technical one.
Open-source releases lower the cost of access for developers and researchers, particularly those working outside the small set of well-funded Western labs. They also complicate efforts to control how powerful AI systems are deployed, since once weights are released they cannot be recalled. The result has been an unusually fluid competitive environment, with leadership shifting between labs on a quarterly basis. Our technology coverage has tracked the broader competitive landscape over the past year.
According to the Stanford AI Index, Anthropic currently leads the top model rankings, with xAI, Google, and OpenAI close behind. As of March 2026, the gap between the leading US model and DeepSeek's strongest model was reported at less than three percentage points on standard evaluations.
What Is Actually Happening
The V4 series, released on Hugging Face in late April, comprises two preview models. V4 Flash is positioned as a faster, lower-cost option suitable for high-volume tasks, while V4 Pro is the flagship model intended to compete on the most demanding reasoning and coding benchmarks. DeepSeek says the new models incorporate architectural changes and optimisation improvements over the V3 generation.
The most significant technical claim is around what the company calls Hybrid Attention Architecture, a method designed to improve the model's ability to track context across very long conversations. The launch also pushes the supported context window to one million tokens, a leap that allows entire codebases or full-length documents to be processed as a single prompt. Industry analysts note that one million tokens is a notable threshold and one only a handful of frontier labs have reached.
On benchmarks released alongside the launch, DeepSeek claims top-tier performance in coding tasks and significant gains in agentic capabilities, the loose category of tests measuring whether a model can plan and execute multi-step actions on behalf of a user. Independent verification of the figures is still in progress, but early third-party evaluations broadly support the company's claims for the coding benchmarks.
The V4 launch lands during what industry observers have dubbed the "earnings superweek", with Alphabet, Microsoft, Meta, and Amazon reporting quarterly results in close succession. Meta in particular drew investor attention by raising its 2026 capital expenditure forecast to as high as 145 billion US dollars, much of it earmarked for AI infrastructure.
Competing Perspectives
Reactions to the V4 release fall into several camps. Open-source advocates have welcomed the launch as evidence that the gap between proprietary and open systems continues to narrow, and argue that broader access to capable models is crucial for research, academic work, and developers in regions without ready access to commercial APIs. They point to Singapore, the United Arab Emirates, and several other markets where open Chinese models have become the default foundation for local AI development.
Inside US labs, the response has been more guarded. Several senior researchers at American firms have argued that open-source releases of frontier-capability models accelerate proliferation risks and complicate efforts to build in safety controls. Others see the competitive pressure as a useful corrective to what they describe as a tendency by large American labs to prioritise commercial deployment over public benefit.
Policy circles in Washington have been particularly attentive. The US Commerce Department recently ordered chip equipment companies to halt certain shipments to Huahong, China's second-largest chipmaker, in the latest round of export restrictions aimed at slowing Chinese semiconductor development. Whether such measures will meaningfully constrain the underlying AI capabilities is increasingly questioned, given that DeepSeek's progress has been achieved despite existing restrictions. For wider context, see our business analysis on US-China trade.
The Alverno Alpha Analysis
The conventional framing of the AI race as a binary US-China contest is starting to look outdated. What V4 actually demonstrates is that the structure of the global AI landscape is more layered. There are still only a handful of labs at the absolute frontier, but the open-source ecosystem now produces models that are within a few percentage points of those frontiers, and that ecosystem is increasingly anchored in China. For most practical applications, the difference between a top-five model and a top-one model is not large enough to drive deployment choices.
That has consequences for the export-control playbook that has dominated US policy for the past three years. Restrictions on advanced chips were predicated on the assumption that capability gaps would widen over time and that controlling the most advanced hardware would lock in a durable lead. The opposite is happening. Chinese labs are achieving more with less, and the open distribution of their results means the geographic location of the lab is becoming less relevant to where the technology actually gets used.
Worth watching is how the major US labs respond. There has been a clear trend among them to disclose less about training data, parameter counts, and architecture decisions. That trend is likely to accelerate, but it sits uncomfortably with the public-interest framing many of those labs have built their public messaging around. A strategy of secrecy in response to open competition is defensible, but it is going to be harder to defend rhetorically than it has been in the past.
Key dates ahead include the formal release of the V4 production models, expected in the coming weeks, and Apple's quarterly results. Apple's on-device AI strategy, built on the M5 silicon generation, is being closely watched as a possible counter-narrative to the cloud-and-scale approach pursued by most of its peers.
Conversation 0 comments
Join the discussion ↓