Our Blog

Blog Index 

DeepSeek and MiniMax Clear China's AI Export Filings: What the New Rules Actually Allow

Posted on 23rd Oct 2026 12:03:50 in Artificial Intelligence, Machine Learning

Tagged as: DeepSeek, MiniMax, China AI Export Rules, Open-Weight AI, AI Regulation

China's two most closely watched open-weight AI developers have walked through the country's new export paperwork and come out the other side with their businesses intact. On the evening of October 22, the CCTV-affiliated account Yuyuan Tantian reported that DeepSeek and MiniMax had completed export registration procedures and had proceeded with open-source releases and overseas licensing.

The disclosure landed six days after Beijing published a draft revision of its catalogue of technologies that are banned or restricted from export. That draft added 23 AI-related items covering large language models, AI agents, computing clusters and information processing, and it came with a registration requirement that took effect immediately, before the consultation period had even closed.

For anyone outside China building products on top of DeepSeek, MiniMax, Qwen, Kimi or GLM, the practical question is simple: does this close the tap on cheap Chinese models, or does it formalise a pipeline that was already running? The short answer, based on the draft text and the commerce ministry's own explanations, is the second one, with new paperwork attached to the most sensitive transfers.

What DeepSeek and MiniMax Actually Completed

The registration is not a licence to operate and it is not an approval of a specific model. It is a filing obligation. Under the temporary rule that took effect on October 16, companies must register certain cross-border transactions with the Ministry of Commerce before proceeding. The covered categories include transfers of technology ownership, overseas licensing arrangements, changes in ownership or control that place a Chinese technology holder under a foreign entity, and transactions that have already begun but are not yet complete.

The filing asks for the exporter, the recipient, the destination, the technology involved, its intended use and the expected value of the export. Regulators want visibility into who is moving what across the border while the final catalogue is still being written.

What DeepSeek and MiniMax did, according to the state media account, was complete that registration and then continue with open-source releases and overseas licensing. Yuyuan Tantian framed the measures as an attempt to draw clearer regulatory boundaries rather than shut the door. "Simply put, AI companies can continue with their commercialisation efforts as usual," the account said. "Once the relevant measures are officially implemented, the vast majority of artificial intelligence technologies will still be freely exportable." It also urged companies to "remain calm in the face of the new regulations."

That is state media talking, not an independent regulator, and the draft is still open for public comment until November 15. But the sequence matters: two flagship developers tested the process in public and were allowed to keep shipping.

The Rules Behind the Filing: What Is Restricted and What Is Prohibited

The draft revision covers 40 technologies in total across AI, biomedicine and solar power. Fourteen are proposed for prohibition and 26 for restriction. The AI bucket is where most of the international attention has gone, because it touches model weights, source code and chip design rather than finished consumer products.

  • Prohibited: core technology for large models trained on sensitive national data, AI agents designed for highly sensitive environments such as national defence and critical information infrastructure, and systems that can autonomously generate or modify malicious code.
  • Restricted: model weights and source code for systems that meet certain capability thresholds, AI-assisted vulnerability discovery tools, advanced agent technology, computing clusters, and chip design for AI processors and memory chips built on 14-nanometre or more advanced processes or using 2.5D and 3D packaging.
  • The size trigger: models with at least 2 trillion parameters, or that rank among the world's top 10 on performance, usage or safety measures, can fall into the restricted category. The draft does not name individual commercial models, though Moonshot's Kimi K3 and Alibaba's Qwen 3.8-Max are among those publicly reported to exceed the parameter threshold.

Those thresholds explain why the announcement caused a stir. Several of the models that Indian developers and startups actually use sit near or above the line, and a strict reading could have turned routine self-hosting into an export event. The commerce ministry's accompanying policy note is what defuses that reading.

Why Open Weights Survive the Crackdown

The draft draws a distinction that is easy to miss and important to understand: transferring controlled technology is not the same as giving foreign customers access to a service built on it.

According to the Ministry of Commerce, technologies that have already entered the public domain through open-source releases or academic papers are proposed to fall outside the scope of export management. AI open-sourcing, overseas data centres built and operated by Chinese companies, joint research using overseas resources, chip fabrication abroad and participation in international standards work would qualify for simplified or automatic licensing, which the ministry describes as close to export upon filing.

On the access side, serving a restricted model through API calls, websites, apps, software-as-a-service products or lightweight on-device deployment does not count as exporting large-model technology. Local deployment of a restricted model overseas does fall under export management, although running it on a company's own overseas cloud can use the automatic licensing route.

In plain terms, the rules target the transfer of the underlying recipe, not the sale of the dish. A Bengaluru startup calling DeepSeek through an API is in a different category from a company that takes the weights and stands up its own inference cluster in a foreign data centre, and only the second case now involves a filing step.

The Timing Was Not a Coincidence

Both companies had open-source news ready in the same window. Hangzhou-based DeepSeek upgraded its V4.1 Flash model and, on October 20, open-sourced its large-scale inference framework DeepInfer along with a companion toolkit called deepseek-recipe. Shanghai-based MiniMax published the M3.1 model on October 23, days after putting M3.1 Flash on its API.

MiniMax also did something more directly relevant to foreign teams. It set up a formal application route for overseas regions that its earlier model licence had excluded, specifically the United States, the European Union, the United Kingdom and South Korea. The weights stay open, but organisations in those regions need to pass a compliance review and obtain a licence before deploying the models locally. India is not on that list, which means Indian teams using the API or the open weights are not pulled into that particular review, though licence terms are still worth reading before any commercial self-hosting.

Read together, the filings and the releases look less like a retreat from open source and more like a licensing regime bolted onto it: keep publishing weights, keep serving the world through APIs, but register the transfers that Beijing considers strategically sensitive.

What This Means for Teams Building on Chinese Models

For most businesses outside China, nothing changes this week. API access is explicitly carved out of the draft, and open weights that are already public are proposed to sit outside export management. The cost advantage that made these models attractive in the first place remains intact.

The changes that do matter are about diligence rather than availability:

  • Check the licence, not just the benchmark. Open weights are not the same as unrestricted commercial use, and MiniMax's regional application route shows how quickly licence terms can diverge by geography.
  • Separate API use from self-hosting. If your product calls a hosted endpoint, you are buying a service. If you download weights and run them in your own infrastructure, you are closer to the transfer that regulators now want to see filed, especially if that infrastructure sits in a third country.
  • Keep a second provider warm. The final catalogue could change after the November 15 comment deadline. Any agent pipeline that depends on a single Chinese model should have a tested fallback, if only for procurement reasons.
  • Watch the chip-design clauses. The draft also touches design technology for AI processors and advanced memory chips. That is aimed at the semiconductor supply chain, but it feeds directly into the cost and availability of the hardware that inference depends on.

The Reliability Caveat That Comes With the Discount

Cheap tokens are only cheap if you do not have to redo the work. Independent testing of MiniMax's M3.1 Flash is a useful reminder of that.

Artificial Analysis scored M3.1 Flash at 42 on its Intelligence Index, up 13 points from the previous M3, at roughly $0.22 per task against about $0.51 before. Humanity's Last Exam came in at 48 percent and the physics benchmark CritPt at 22 percent. On price and raw capability that is a genuine step forward, and it puts the model level with Z.ai's GLM-5.3 Flash, just ahead of Google's Gemini 3.8 Flash and behind DeepSeek's V4.1 Flash-1017 at 45.

The same evaluation found something less comfortable. On the AA-Omniscience factual test, M3.1 Flash answered 56 percent of questions correctly, but it attempted almost every question rather than declining when unsure, producing a 93 percent hallucination rate under the benchmark's definition. That figure does not mean 93 percent of its answers were wrong. It means that when the model did not know, it overwhelmingly chose to invent an answer rather than admit uncertainty, which dragged its overall score on that test down to 16.

For a shop assistant drafting product descriptions, that is an annoyance. For an agent reconciling invoices, filing compliance paperwork or summarising a contract, it is a design constraint. The fix is architectural rather than rhetorical: give the model the source text, require a verification step, and never let it be the only thing standing between a number and a customer.

The Bigger Picture: Export Controls, Both Directions

China's move did not happen in a vacuum. It arrived in the same week that Washington and Beijing reported progress in AI safety talks and agreed to set up a notification mechanism for AI incidents that could threaten national security. It also follows years of US-led restrictions that have kept the most advanced chips and chipmaking equipment out of China.

"The move can be seen as a response to Western countries posting sovereignty requirements on AI, robotics, telecommunications, and other key technologies," Lian Jye Su, chief analyst at Omdia, told the Associated Press. Gary Ng, a senior economist at Natixis, put the broader trend more bluntly: "The world is getting more fragmented in trade and investment, but also in access to technologies."

The symmetry is worth noting. In June, the US Commerce Department asked Anthropic to bar foreign nationals from using some of its most powerful models, citing national security; those restrictions were later lifted. Both governments now treat frontier model access as a strategic lever, and both are building filing systems that distinguish between commercial service and technology transfer. For companies in the middle, the compliance surface is expanding in every direction.

What to Watch Before November 15

The draft catalogue is a consultation document, and the final scope can still move. Three things are worth tracking: whether the 2-trillion-parameter threshold survives intact, whether the open-source carve-out is narrowed for models trained on sensitive data, and whether the registration requirement becomes a standing obligation after the final revision takes effect.

For now, the signal from Beijing is that open-weight releases and overseas API service are still part of the strategy. DeepSeek and MiniMax have shown the process can be completed. The next test will be whether smaller Chinese labs, and the foreign companies that license from them, find the same path as easy to walk.

Sources

Recent Posts

Catgories

Archives

`
whatsapp me