DeepSeek keeps landing in the middle of one argument. Should powerful AI models ship with open weights, or stay behind a paid API? A Chinese Embassy piece titled The wisdom of openness: What DeepSeek says about China’s vision describes DeepSeek as open source, built by locally educated Chinese engineers for a stated $6 million. That framing is the clearest official-adjacent answer to why Chinese AI leans open, and it frames DeepSeek as a counterweight to OpenAI’s ChatGPT.
Others see the same openness and hesitate. Researchers reviewed for this story caution that safeguards can be stripped out of open-weight models, even though the same research says that removal does not necessarily lead to greater real-world harm. Both halves of that finding matter, and this analysis keeps them side by side.
The commercial stakes are rising too. Our earlier report on DeepSeek’s reported $15 billion raise ahead of an IPO showed how much capital now rides on the idea that a lab can publish its models and still build a business. This article, dated October 26, 2026, separates what the available sources confirm from what commentators merely assume.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the Sources Confirm About DeepSeek Open Source
The confirmed record is short, so it pays to list it plainly. The Chinese Embassy article calls DeepSeek open source. It says local engineers educated in China built the model. It puts the development cost at $6 million. It says the team made its source code and training data transparent, and it sets the model against OpenAI’s ChatGPT. It also says DeepSeek aims to empower businesses and accelerate innovation.
Every one of those points comes from a single embassy-hosted article. That matters for how much weight readers should give them. An embassy has a clear interest in presenting a flagship national AI project as transparent, frugal and homegrown. The claims may well be accurate, but this article treats them as the sponsor’s account rather than an audited record.
The second confirmed strand comes from the research side. Open-weight models, DeepSeek’s included, can have their safeguards removed. The sources name no specific technique and no specific model version. They add that stripping safeguards does not necessarily produce more real-world harm, while researchers still urge caution. That is a hedged finding, and honest coverage keeps it hedged.
The $6 Million Claim and Why It Matters
The $6 million figure carries most of the rhetorical weight in the embassy account. A low development cost supports a story where clever engineering beats brute spending, and it makes openness look like a strategy rather than a charity. If a competitive model costs that little to build, publishing it gives away less than a rival would assume.
Treat the number with care. The sheet of confirmed facts attributes it to the embassy article and offers no breakdown of what the figure covers. Training compute, research salaries, failed experiments and hardware purchases can each be counted or excluded, and the answer swings wildly. No independent audit appears in the sources reviewed here.
Even with that caveat, the figure shapes how the rest of the market reacts. Procurement teams read it as a signal that capable models may get cheaper to run and to adapt. Rival labs read it as pressure on their own pricing. Skeptics read it as marketing. All three reactions help explain why the open-versus-closed question has moved from conference panels to board meetings.
Transparency Pitch: Source Code and Training Data
The embassy article says DeepSeek made both its source code and its training data transparent. In the AI world, that is a bigger promise than publishing weights alone. Weights let users run and adapt a model. Code lets them reproduce the serving stack. Training data lets outsiders check what the model learned from and whether it absorbed anything it should not have.
Whether every layer of that stack is actually available is a separate question from what the article claims. The sources reviewed do not itemize which artifacts exist, and the vocabulary here is slippery. “Open source” and “open weight” often get used interchangeably, although they describe different levels of access. A fair reading is that the model’s weights and code reach the public, with the depth of data disclosure left for outside researchers to verify.
That distinction drives the caution on the other side. A buyer who cannot inspect training data has to trust the vendor’s word on provenance. A buyer who can inspect it still needs the time and expertise to do so. Transparency lowers the barrier to trust, but it does not replace the work of earning it.
Does China Officially Favour Open Models?
The headline question assumes a policy, and the available results do not establish one. No source reviewed here shows the Chinese government formally mandating or even officially preferring open-weight models. What the sources show is narrower. One state-affiliated outlet praises DeepSeek’s openness, and one prominent Chinese lab has chosen to publish.
The South China Morning Post’s AI coverage hints at a more mixed picture. Its AI topic page at the time of writing carries a headline asking whether a new state-owned giant can supply the data fuel China’s AI race needs, and another reporting that a Chinese video model beats US rivals on visual quality but lags on safeguards. Neither headline describes a blanket open-model mandate. They describe a sector with state involvement, competitive pressure and uneven safety practice.
So the more accurate framing is behavioral rather than legal. Chinese labs, led by DeepSeek, have published models openly, and Chinese official messaging has celebrated that choice. Whether Beijing plans it, merely welcomes it, or lets individual companies decide is something the sources leave open.
Why Open Weights Suit Chinese Labs: Four Hypotheses
Commentators offer several explanations for the Chinese tilt toward openness. None of them is confirmed in the sources reviewed, and the fact sheet for this story marks each as unconfirmed. They are worth laying out because they frame what a reader should watch for next.
Sovereignty and Supplier Risk
The first argument is dependence. A firm that runs open weights on its own servers does not rely on a foreign API provider, a foreign cloud contract or a foreign pricing decision. For organizations with data-residency rules, local deployment also keeps sensitive records inside the building. This is the most intuitive explanation, but the sources do not tie it to any named Chinese policy.
Cost and Domestic Adoption
The second argument is economic. If a shared open model is cheap to adapt, thousands of smaller companies can build on it without paying recurring access fees. DeepSeek’s own stated aim to empower businesses and accelerate innovation fits that story. Still, claims that openness was chosen specifically to cut costs or speed domestic adoption remain unconfirmed.
A third hypothesis ties openness to chip export limits, and a fourth to geopolitical influence, since widely used models spread a country’s technical standards. The sources reviewed confirm neither. Readers should treat both as speculation until a named official, filing or company statement says otherwise.
Why Others Remain Cautious
The sources confirm that researchers urge caution on open-weight models, but they do not name which governments or companies sit in the cautious camp. The fact sheet is explicit that no formal rejection of DeepSeek on safety, privacy, security or censorship grounds has been confirmed. Any report of a government ban needs its own source.
What can be said is how the cautious position works in practice. Enterprise security teams generally want to audit a model before it touches customer data. Compliance teams want to know who is liable if an output causes harm. Open weights shift responsibility toward the deployer, because no vendor stands behind a model that anyone can modify. That shift makes some buyers uncomfortable even when the model is free.
The geopolitical overlay adds friction. Our report on 29 countries signing WAICO as the rival Pax Silica bloc takes shape shows governments being asked to pick sides on AI infrastructure. A model’s country of origin therefore enters procurement decisions alongside its benchmark scores, whether or not the model itself is technically sound.
Safeguards in Open-Weight Models: The Harm Debate
The research thread in the available results deserves a closer look because it cuts against both easy narratives. On one side, DeepSeek’s safeguards can be removed once weights are public. Anyone with the files can fine-tune away refusals. That fact alarms safety researchers, because a closed API can revoke access and an open release cannot be recalled.
On the other side, the same results state that removing safeguards does not necessarily lead to greater real-world harm. A model that answers more freely is not automatically a model that enables more damage, since real harm usually needs resources, intent and skill beyond a chatbot’s answers. Researchers still caution, which tells us the question is unsettled rather than resolved in either direction.
The policy angle sits close to this. Manhattan Institute researchers have argued for a legal approach that protects open-weight releases, which we covered in our look at a safe harbor for open weights framed against China. That debate shows the caution is not limited to one side of the Pacific. Western policymakers are weighing the same trade-off about their own open releases.
Voices From the Industry on Open Release
Four named statements, all drawn from published sources, show how the openness argument sounds from different seats. They date from the period after DeepSeek’s R1 release and from the founder’s own explanations, so they speak to the principle more than to the latest model.
Yann LeCun, chief AI scientist at Meta, told NBC News that the story is not about China surpassing the US. In his words, it is that “open source models are surpassing proprietary ones.” That reading moves the debate from nationality to licensing model.
Marc Andreessen, co-founder of Andreessen Horowitz, called DeepSeek R1 “one of the most amazing and impressive breakthroughs I’ve ever seen” and described it, as open source, as “a profound gift to the world.” A prominent US investor praising a Chinese release shows that the cautious camp is not monolithic.
DeepSeek founder Liang Wenfeng supplied the lab’s own rationale, reported in a Cato Institute blog post: “We won’t go closed-source. We believe that establishing a robust technology ecosystem matters more.” Liang frames openness as ecosystem-building rather than as an act of national policy, which fits the point that no government mandate has been confirmed.
The policy newsletter ChinaTalk put the upside simply: “DeepSeek’s greatest strength lies in its open-source approach, which empowers researchers worldwide.” Taken together, the four statements make the case for openness. The caution described elsewhere in this article comes from researchers and buyers rather than from these speakers.
DeepSeek vs ChatGPT: Open and Closed Compared
The embassy article sets DeepSeek against OpenAI’s ChatGPT, and that contrast is the cleanest way to see what each approach buys. The table below is an analytical framework built from the confirmed points and general industry practice. It contains no benchmark measurements, since the sources reviewed supply none for a head-to-head test.
| Dimension | Open-weight model (DeepSeek-style) | Proprietary API (ChatGPT-style) |
|---|---|---|
| Weight access | Public files that anyone can download and adapt | Hosted only, no access to weights |
| Cost structure | Hardware and engineering cost, no per-call fee to the lab | Per-use fees set by the vendor |
| Data residency | Can run inside the buyer’s own environment | Prompts travel to the vendor’s servers |
| Auditability | Inspectable by outside researchers | Depends on what the vendor chooses to disclose |
| Safeguard control | Can be removed by anyone with the weights | Enforced by the vendor at the API layer |
| Vendor accountability | Limited, responsibility shifts to the deployer | Clearer, a single company answers for the service |
| Stated origin story | Embassy article cites $6 million and local engineers | No comparable cost figure in the sources reviewed |
Our own DeepSeek vs Claude pricing comparison covers how those cost structures play out per million tokens. The short version is that the per-call fee is the visible cost, while integration, hosting and governance are the hidden ones.
Market Impact: Pricing Pressure and Developer Choice
An openly published model with a stated $6 million build cost pushes on the whole market in three ways. It gives developers a free baseline, so any paid model must justify its premium with quality, reliability or support. It gives enterprises a credible alternative to negotiate against. And it lets smaller teams ship products without a vendor relationship at all.
Open competition is not limited to one lab. Our coverage of the Kimi K3.1 Pro open-weight debut from Moonshot AI shows that the Chinese open-weight field has more than one serious entrant. That depth matters for the “why China” question, because a pattern across several labs points to shared incentives rather than one founder’s taste.
The downside risk sits with buyers who adopt too quickly. Tom’s Hardware reported that DeepSeek’s latest model brings low running costs but at a price, with the open-weight model suited to agentic workloads yet more prone to hallucination than proprietary competitors. Cheap inference helps only if outputs are trustworthy enough to act on, and that trade-off is where many pilots will succeed or stall.
Historical Context: Open Source Has Won Before
The current fight echoes older ones in software. Linux and Android both grew into dominant platforms by letting many parties build on a shared base, while commercial vendors competed on services layered above it. Supporters of open AI make the same argument today. A shared model becomes infrastructure, and the money moves to applications, hosting and tuning.
The analogy has limits that the safety debate exposes. Operating systems do not generate persuasive text or write exploit code on request. An open AI model‘s behavior depends on training, and the person holding the weights can retrain it. That gives the caution in the sources a basis the Linux era never had, even if the commercial logic looks familiar.
LeCun’s remark that DeepSeek’s team built on other people’s published work captures the continuity. Open research compounds across borders, and a lab that publishes also gets to build on what others publish. That cycle favors whoever joins it early, which is one plausible reason labs choose openness regardless of their government’s stance.
The Latest DeepSeek Update and What It Signals
DeepSeek has kept shipping. Interconnects, a newsletter that tracks open model releases, titled a recent issue around the DeepSeek V4.1 Flash 1017 release alongside American-made models and new regulatory controls in the US and China. The pairing in that title is telling. Releases, rival open models and regulation now move as one story.
For readers choosing a model to run locally, the practical comparison lives in our best local LLM roundup featuring DeepSeek V4.1 Flash. This analysis does not restate benchmark scores for the latest update, because the fact sheet for this story does not confirm them. SiliconANGLE and Tom’s Hardware both covered the update, and readers who need figures should consult those reports directly.
The cadence itself carries a message. A lab that ships frequent revisions in public invites scrutiny and feedback, and that feedback can improve the product faster than closed testing. It also means any evaluation goes stale quickly. A security review of one version may not hold for the next.
Claim Ledger: Confirmed vs Unconfirmed
The table below sorts the claims around this story by how well the available sources support them. It exists so readers can quote the confirmed items without accidentally repeating the speculative ones.
| Claim | Status | Basis |
|---|---|---|
| DeepSeek is open source | Stated | Chinese Embassy article |
| Built by locally educated Chinese engineers | Stated | Chinese Embassy article |
| Development cost of $6 million | Stated, not independently audited | Chinese Embassy article |
| Source code and training data made transparent | Stated, depth unverified | Chinese Embassy article |
| Safeguards can be removed from open-weight models | Confirmed in research reviewed, no method named | Research results reviewed |
| Removal does not necessarily raise real-world harm | Reported finding, researchers still urge caution | Research results reviewed |
| China’s government officially favours open models | Not established | No source reviewed |
| Openness chosen to bypass US chip limits | Unconfirmed | Commentary only |
| Foreign governments or firms formally rejected DeepSeek | Unconfirmed | No source reviewed |
Five Predictions for the Open-Model Debate
These are analyst judgments, not reported facts. Each follows from the dynamics described above, and each could prove wrong.
- Procurement checklists will formalize. Enterprises that adopt open-weight models will add provenance, audit and fine-tuning-control requirements to vendor questionnaires.
- More non-Chinese labs will publish weights. Interconnects already tracks American-made open models, and competitive pressure from free baselines tends to widen that field.
- Safeguard-removal research will draw regulators. The same Interconnects issue links releases to new regulatory controls in the US and China, so expect rules to target how weights are released and modified.
- Price pressure will persist. A stated $6 million build cost gives buyers leverage, and proprietary vendors will answer with bundled services rather than raw price cuts.
- Deployment will split by region and sensitivity. Chinese firms will favor local open weights, while many foreign firms will run hybrids, using open models for low-risk work and audited vendors for sensitive tasks.
For ongoing coverage of model releases and policy, the AI and machine learning section tracks each development as it lands.
What Readers and Builders Should Do Now
Developers weighing DeepSeek’s open models should test them on their own tasks instead of trusting any headline number, including the $6 million figure. Pin the exact model version in use, because the lab revises often. Log outputs for review, and keep a human check on any agentic workflow where a hallucination could cause real cost.
Security and legal teams should ask the questions the cautious camp raises. Where do the weights come from, and can the hash be verified? Who maintains the deployment, and who answers if it fails? Which data may enter the model? Open weights make these questions easier to answer than a closed API does, but only when someone actually does the work.
Policy readers should hold two ideas at once. Open release has delivered real benefits, as LeCun, Andreessen and Liang each argue. It also hands every downstream user the power to remove safeguards, which is why researchers stay careful. The honest position is that the evidence supports neither triumph nor panic.
FAQ
Does China officially favour open AI models?
The sources reviewed do not establish an official policy. They show that DeepSeek publishes openly and that a Chinese Embassy article praises that openness. A formal mandate or preference is unconfirmed.
Is DeepSeek really open source?
A Chinese Embassy article describes it that way and says the team made its source code and training data transparent. Open source and open weight are different things, so outside researchers still need to verify how much is actually released.
How much did DeepSeek cost to build?
The embassy article states $6 million. That figure has not been independently audited in the sources reviewed, and it does not say what costs it includes.
Can safeguards be removed from DeepSeek models?
Yes, the research reviewed says safeguards can be removed in open-weight models. It names no specific method or version. The same research says removal does not necessarily lead to greater real-world harm, although researchers urge caution.
Which governments or companies have rejected DeepSeek?
None is confirmed in the available sources. Claims of formal rejection over safety, privacy, security or censorship need their own named source before anyone should repeat them.
Why do open-weight models worry some buyers?
Open weights shift accountability to whoever deploys the model, and they let anyone modify its behavior. Buyers with strict audit, liability or data-handling rules often prefer a vendor that answers for the service.
What is the newest DeepSeek release?
Interconnects covered a DeepSeek V4.1 Flash 1017 release in a recent issue, and SiliconANGLE and Tom’s Hardware reported on the update. Check those reports and DeepSeek’s own channels for current details, since the lab revises its models often.
