DeepSeek releases latest AI model reshaping China-US race
Updated
Updated · The New York Times · May 3
DeepSeek releases latest AI model reshaping China-US race
14 articles · Updated · The New York Times · May 3
The Chinese start-up unveiled the model last week, after its early-2025 system became the most downloaded free app in the United States.
The new release is portrayed as pushing China and the US into a neck-and-neck contest while changing competition from sheer scale and spending to efficiency.
DeepSeek's earlier low-cost breakthrough challenged assumptions that China lagged in AI, prompting comparisons in Silicon Valley to a technological Sputnik moment.
Is China’s low-cost AI a true innovation or a strategy based on illicitly 'distilling' Western technology?
Did U.S. chip sanctions backfire, accidentally fueling China's rise as a leader in artificial intelligence?
As AI’s power demand soars, will the race be won by better code or by a superior energy grid?
DeepSeek V4 Launch: 1 Million-Token Context and Cost Revolution Reshaping the Global AI Race
Overview
In May 2026, DeepSeek AI launched DeepSeek V4, featuring two variants: V4-Pro for complex reasoning and coding, and V4-Flash for fast, cost-efficient tasks. Its breakthrough 1 million-token context window, combined with innovations like compressed sparse attention, Mixture-of-Experts architecture, low-precision inference, and the Engram memory system, enabled top performance and remarkable cost savings. These efficiencies allowed aggressive pricing that disrupted the AI market and expanded global access, especially through an open-source MIT license. Despite privacy bans in several countries and geopolitical tensions fueled by US export controls and IP theft allegations, DeepSeek’s strategy exemplifies China’s focus on efficient, agentic AI deployment and ecosystem openness, challenging US dominance and reshaping the global AI landscape.