The US Commerce Department quietly withdrew a draft regulation this month that would have required export permits for AI chips shipped to any country in the world, including close allies like the UK, Japan, and the Netherlands. Around the same time, the same administration began approving limited exports of higher-tier chips to China under specific licenses. After years of tightening restrictions, Washington reversed course — not because the underlying concern about Chinese AI capability had changed, but because the rule had become politically untenable with trading partners.
Chinese labs, meanwhile, have continued apace.
On March 11, a model calling itself Hunter Alpha appeared on OpenRouter, an AI routing platform that aggregates access to models from dozens of providers and allows developers to call any of them through a single API. OpenRouter requires no public disclosure from the labs that list models on it, which means a sophisticated AI system can be made globally accessible with no announcement, no press release, and no attributed developer. Hunter Alpha arrived exactly this way. Describing itself as "a Chinese AI model primarily trained in Chinese," it listed a knowledge cutoff matching DeepSeek's own chatbot and declined to identify its creators.
The specifications have drawn immediate attention. Hunter Alpha claims 1 trillion parameters and a 1 million token context window. It describes its purpose not as general conversation but as a "heavy engine for agentic tasks" — long-chain planning, complex reasoning, and multi-step tool execution. Early developer testing finds it reliable on structured agentic workflows and above-average on tool-calling, though weak on creative tasks and mathematics. Wharton professor Ethan Mollick, who ran it through logical reasoning benchmarks, called it "okay."
The leading theory is that Hunter Alpha is DeepSeek testing V4 ahead of a planned April launch: the parameter count and context window match specifications that have circulated since early this year, a "V4 Lite" variant briefly appeared on DeepSeek's own website days before Hunter Alpha went live, and its chain-of-thought reasoning pattern closely resembles DeepSeek's prior models. A second theory points to Zhipu AI: the same anonymous OpenRouter account previously published a model later identified as Zhipu's GLM-5. Neither lab has commented.
The episode follows a pattern that has repeatedly caught Western policymakers off-guard. When DeepSeek released its R1 reasoning model in January 2025, trained for a reported six million dollars, the announcement wiped roughly 593 billion dollars from NVIDIA's market capitalization in a single trading session, the largest single-day loss in market history for any company. The disruption was not military or geopolitical: it was a demonstration that the compute-cost assumptions underlying the Western AI industry might be wrong. That result was achieved under the chip export restrictions then in force.
Washington's difficulty is not just that the rules are hard to enforce. It is that the policy question, whether restricting hardware access meaningfully limits Chinese AI capability, does not have a settled answer, and the evidence arriving from OpenRouter this month does not resolve it in the direction policymakers would prefer. Hunter Alpha is free to any developer with an internet connection.
Sources: Reuters, Bloomberg, Reuters
--
By the Control Plane Editorial Team