China Green-Lit DeepSeek and MiniMax to Take Their AI Global
Beijing spent last week scaring the open-source AI world. This week it turned around and handed two of its champions a licence to keep flooding that world. If that sounds like a contradiction, you are reading the wrong part of the story — the part Beijing wants you to read is the one where the scariest export-control regime in Chinese tech history turns out to be a velvet rope, not a wall.
DeepSeek and MiniMax have completed the new registration procedures and are proceeding with open-source releases and overseas licensing of their AI models, according to Yuyuan Tantian, a social-media account affiliated with state broadcaster CCTV, which broke the news in a Thursday-evening post. In a follow-up post on Friday, the account said overseas internet users had urged other Chinese open-source large language models to follow suit. DeepSeek and MiniMax did not immediately reply to requests for comment, Reuters reported.
We pulled the underlying reports ourselves — and one small detail fell apart on inspection: Reuters carried MiniMax with the tag (0100.HK), which is Tencent's Hong Kong listing, not MiniMax's; MiniMax is privately held. We dropped the ticker rather than print it.
What Beijing actually announced — and didn't
Read the announcement carefully, because Beijing certainly did. Nobody issued a ministry press release. There was no commerce-ministry decree with a red seal granting DeepSeek and MiniMax the right to export their models. What happened, per both Reuters and the South China Morning Post, is that the two companies completed the registrations required under the new temporary rules, and a state-media-affiliated account presented that as proof the regime is permissive. The medium is part of the message: Yuyuan Tantian has become Beijing's unofficial loudspeaker for precisely this kind of reassurance — official enough to be authoritative, informal enough to be deniable.
The substance behind the signalling is real enough. In recent days, Hangzhou-based DeepSeek upgraded its V4.1 Flash model and open-sourced its in-house inference framework, while Shanghai-based MiniMax open-sourced its M3.1 model on Friday. Both moves happened under the new rules, and both were allowed through — which is exactly the point Beijing wanted made. The post urged businesses to "remain calm in the face of the new regulations," a line that reads less like a regulator's notice and more like a pep talk.
The export controls that caused the panic
The panic was understandable. On October 16, the Ministry of Commerce and the Ministry of Science and Technology jointly published a draft revision of the Catalogue of Technologies Prohibited or Restricted from Export — Announcement No. 48 — proposing to add 40 technologies, 23 of them in artificial intelligence, to China's export-control list. The AI items cover large language models, AI agents and computing clusters, and one provision took immediate effect: any transfer of covered technology, including overseas licensing deals and changes in foreign ownership or control, must now be registered with the ministry before proceeding. Deals already underway are covered too. The draft is open for public comment until November 15.
The net is deliberately wide. A model is caught if it meets any one of three triggers: at least 1026 training FLOPs, at least two trillion parameters, or a top-10 global ranking. Also restricted: advanced agents with multi-step planning and tool use, AI chip designs at 14 nanometres or below, computing clusters of 10,000 or more accelerators, and cross-vendor chip scheduling. But the same announcement made clear what is not covered: serving overseas users through APIs, apps, websites or SaaS, and any technology already released as open source or in academic publications. For open-sourcing, a commerce-ministry spokesperson promised expedited "automatic licensing" — as quickly as "immediate reporting and export." Most AI technologies, the ministry said, "will still be freely exportable."
"Simply put, AI companies can continue with their commercialisation efforts as usual," Yuyuan Tantian said. "Once the relevant measures are officially implemented, the vast majority of artificial intelligence technologies will still be freely exportable."
The sharper read comes from outside the state media. "The restrictions seem to be designed more as a switch than a wall," wrote the Takshashila Institution in its item-by-item analysis of the order. "Almost nothing changes for exporters today, but China now holds the means to monitor, slow or stop the flow of its AI technology abroad whenever it chooses." Automatic licensing is still licensing — open release has moved from permitted by default to permitted under a licence, and a licence can be narrowed or withdrawn. Beijing has been clear about its own logic: the Yuyuan Tantian post said "Our revision is a routine process carried out in accordance with laws and regulations," while insisting China "is not taking the path" of the United States in using such measures to isolate competitors and preserve its technological lead.
What the approval actually covers: registration, not a blank cheque
Read the Yuyuan Tantian wording closely and this is narrower — and more interesting — than "China approves DeepSeek's global release." The account's Thursday post said DeepSeek and MiniMax had "successfully completed the registration process, as well as the procedures for open-sourcing and authorising overseas use," and added that "AI companies can continue their commercialisation efforts as usual." That's a green light for business as usual under a new registration regime, not a sweeping declaration that anything Chinese labs build can now go anywhere.
A follow-up post on Friday pushed the signal further: overseas internet users had been urging other Chinese open-source models to follow suit, positioning DeepSeek and MiniMax as the first wave rather than a one-off exception. Beijing has run this playbook before — it added algorithms to its export-control catalogue in 2020, right after the first Trump administration threatened to ban TikTok unless it sold its US assets. The reporting for both the registration wording and the 2020 precedent comes via Bloomberg's relay of The Edge Singapore and The Edge Malaysia (Oct 23, 2026) and Reuters (Oct 23, 2026). And the usual caveat still holds: the approval was announced through a state-media-affiliated social account, not a ministry decree, and the export-control regime itself is still in draft — so treat "registration" as a sanctioned signal, not final statutory clearance.
Across the Pacific, Washington is going the other way
The asymmetry is the story. While Beijing was reassuring the world its open models would keep flowing, Washington was bolting its own doors. On October 20, the Department of Homeland Security published Exclusion and Removal Order #2026-01 — signed by Secretary Markwayne Mullin on July 24 — directing every US federal civilian agency to rip out DeepSeek products and third-party tools "developed using DeepSeek implementations" from their IT systems, and banning any future purchases from DeepSeek or its affiliates. The order said DeepSeek's models pose "multiple threat vectors" and warned that their use "may result in the transfer of sensitive data to the PRC" by US government employees or vendors. Beijing called the move discriminatory.
Days earlier, the Treasury's new outbound-investment rule had formally banned US capital from flowing into Chinese AI — a final rule that landed the week of October 19. So the picture, as of this morning: Washington is barring Chinese models from its own government and barring American money from Chinese AI, while Beijing is fast-tracking its best open models into everyone else's hands. The two governments agree on exactly one thing: the weights are strategic. They disagree on whether strategic means locked down or handed out.
The Manus precedent explains everything
In December 2025 Meta closed a $2 billion acquisition of Manus, the Chinese-born AI agent start-up; on April 27, 2026, China's National Development and Reform Commission ordered the deal unwound — the first confirmed use of China's foreign-investment security review to reverse a completed cross-border AI transaction. Manus was forced back into independence, then raised $500M+ in October at a valuation reportedly around $4 billion. "The immediate task for Manus now is proving scale, profitability and regulatory alignment," said Han Lin, China country director at The Asia Group, in CNBC's report.
The pattern is now unmistakable. Beijing will not let a frontier Chinese AI company fall under American ownership — not even when the buyer is Meta. But it will fast-track those same companies' models into the world's hands, because open weights buy what ownership sales don't: developer dependency. That is the logic of the switch. It also helps explain the timing: the registration approvals for DeepSeek and MiniMax landed in the same week as DeepSeek's reported $15B raise, led by CATL and Tencent — a war chest Beijing clearly intends to see spent on global reach, not just domestic training runs.
Our verdict: the decoupling is now two-sided
For years the AI decoupling story was a one-way street: America restricting, China reacting. This week it became two-sided — Washington walls Chinese AI out of its government and its capital markets, while Beijing builds an export-control regime whose chief purpose is to keep Chinese open models flowing into the rest of the world's servers. The October 16 announcement was widely misread as Beijing closing the tap; the real design: monitor everything, license the exports, keep the open weights coming. And here is the uncomfortable part for Washington: Chinese models already run most developer AI traffic precisely because they are cheap, open and good. A licensing switch in Beijing cannot unwind that dependency — but it can, at any moment, tighten it.
Frequently asked questions
Did China approve DeepSeek and MiniMax to export their AI models?
The companies completed the registration procedures required under China's October 16 draft export rules and are proceeding with open-source releases and overseas licensing. The news came via Yuyuan Tantian, a CCTV-affiliated social-media account — not a formal ministry decree — so treat it as a sanctioned signal rather than a legal document.
What do China's new AI export controls actually cover?
The draft (Announcement No. 48) proposes controls on 23 AI technologies: frontier model weights, code and training methods (at 1026 FLOPs, 2 trillion+ parameters, or a global top-10 ranking), advanced AI agents, AI chip designs at 14nm or below, and computing clusters of 10,000+ accelerators. Registration took effect immediately on October 16; public comments close November 15.
Are open-source AI models exempt?
Mostly yes. Technology already released as open source or in academic publications is outside the controls, new open-source releases get expedited "automatic licensing", and serving foreign users via APIs, apps or SaaS is not treated as a technology export.
Why did the US Department of Homeland Security ban DeepSeek?
Exclusion and Removal Order #2026-01 (signed July 24, published October 20) requires federal civilian agencies to remove DeepSeek products and any third-party tools built on DeepSeek implementations, citing "multiple threat vectors" and the risk of sensitive-data transfer to China. It covers civilian agencies only — not the military or the private sector.
What happened with Meta and Manus?
Meta closed a ~$2 billion acquisition of the Chinese-born agent start-up Manus in December 2025; China's NDRC ordered the deal unwound in April 2026. Manus has since resumed independent operations and raised $500M+ in October at a valuation reportedly around $4 billion.
Will these Chinese models keep flowing to developers worldwide?
Yes, for now — Beijing wants this. The longer-term risk is the "switch": the licensing regime gives Beijing a lever to tighten any future release it deems sensitive.
Which Chinese AI companies could be next in line for approval?
Nobody official has named names — and any list that does is guessing. What we know: in a Friday follow-up post, Yuyuan Tantian said overseas internet users had urged other Chinese open-source large-language-model developers to follow DeepSeek and MiniMax through the same registration, and Beijing's line, via the same account, is that "AI companies can continue their commercialisation efforts as usual." Reuters noted that neither DeepSeek nor MiniMax replied to requests for comment on the approval itself. So the queue exists, the door is described as open, and the first wave is set — but the next names through it are still unannounced. Watch the labs with the biggest open-weight footprints.
Sources: Reuters; South China Morning Post; MOFCOM's Announcement No. 48 summaries; onthewire.ai's count and translation of the draft list; the Takshashila Institution's analysis; the Federal Register (DHS Exclusion and Removal Order #2026-01); CNBC; Technode/Pulse2 (via Reuters) on the Manus financing.
By Daniel Rao. Daniel Rao covers the global AI industry for TechPulse USA, reporting on the models, money and regulation shaping artificial intelligence in the US, Europe and Asia.
Comments
Post a Comment