EDITORIAL: DeepSeek V4 Pro under MIT license — poisoned gift or Beijing's masterstroke?

EDITORIAL: DeepSeek V4 Pro under MIT license — poisoned gift or Beijing's masterstroke?

Introduction: Open-source as a geopolitical weapon

April 24, 2026: DeepSeek rewrites the rules

On April 24, 2026, DeepSeek, the artificial intelligence startup based in Hangzhou, published its V4 Pro model on Hugging Face under an MIT license. With 1.6 trillion total parameters (of which 49 billion are active per inference), a context window of one million tokens, and a license that allows unrestricted commercial use, modification, redistribution, and the construction of proprietary products without royalties — DeepSeek V4 Pro became, within weeks, the most powerful open-weight model ever released, and potentially the most disruptive.

For Western developers and businesses, the announcement was a technological gift of the first order. A world-class model, scoring 80.6% on SWE-Bench Verified, ranking first on LiveCodeBench at 93.5%, available for free, downloadable and deployable without legal constraint. At an API cost of $0.435 per million tokens — made permanent since May 22, 2026 — compared to $25 for Claude Opus 4.6, the economic proposition is as revolutionary as the technical performance. The question that no one in Western security circles can avoid: is this really a gift?

The MIT license: the most permissive of open-source licenses

The MIT license is one of the most permissive open-source licenses in existence. It imposes a single obligation: retain the copyright notice in distributions. Everything else — unrestricted commercial use, modification, redistribution, sublicensing, fine-tuning, creation of derivative models, integration into proprietary products — is permitted. It is the same license as React, PostgreSQL, and a large part of the software infrastructure on which the global internet runs.

By choosing MIT rather than Meta's Llama license or a restrictive custom license, DeepSeek sends a clear signal to the global developer community: we want you to use this model without friction. This is a deliberate, strategic choice. It maximizes adoption, builds an ecosystem dependent on the model, and positions DeepSeek as the indispensable open-source alternative to the closed models of OpenAI, Anthropic, and Google. But it also raises questions that the West cannot afford to ignore.

The Western dilemma: performance versus provenance

The concrete advantages for developers and businesses

The advantages of DeepSeek V4 Pro for Western developers are real and difficult to ignore. A context of one million tokens — comparable to Google's best models — at a cost ten to thirty times lower than closed equivalents. Top-tier performance on coding and reasoning benchmarks. A Mixture-of-Experts architecture with 1.6 trillion parameters — the largest of all available open-weight architectures. The ability to deploy the model on your own infrastructure with no per-token cost and no third-party surveillance.

For a startup, an SME, or even a large enterprise with constrained computing budgets, the economic equation is compelling. By comparison, OpenAI's GPT-5.5 costs several dollars per million output tokens. Anthropic's Claude Opus 4.6 reaches $25 per million tokens. DeepSeek V4 Pro API runs at $0.87 per million output tokens, or zero cost if self-hosted. The difference is an order of magnitude. This explains why, according to several analyses, V4 Pro is becoming the default open-weight model for enterprise self-hosted deployment in Q3 2026.

Security risks: the digital Trojan horse

But the question of provenance does not disappear because the license is permissive. DeepSeek is a private Chinese company — founded by Liang Wenfeng, who also runs hedge fund High-Flyer. It operates under the jurisdiction of the Chinese Communist Party, which can legally impose obligations of cooperation with intelligence services. The 2017 National Intelligence Law requires any Chinese organization or citizen to "support, assist, and cooperate with national intelligence work." The model itself — its weights, its architecture — is open-source. But the API service hosted by DeepSeek, which is how most developers first use it, sends data to servers in China.

There is a crucial distinction between the two uses. The self-hosted model — downloading the weights and deploying them on your own infrastructure — involves no data transmission to DeepSeek. The hosted API model, by contrast, exposes user requests and data to Chinese servers. For sensitive applications — medicine, law, defense, finance — this distinction is fundamental. But in practice, the boundary between uses is porous, and developers who start with the API for testing do not always make the transition to self-hosting even for production deployments.

The West caught between open-source ideology and geopolitical reality

The open-source community facing an unprecedented dilemma

The global free software development community has always celebrated open-source as a core value — code belongs to everyone, knowledge must be shared freely. This philosophy built the internet, the cloud, and a large part of the world's digital infrastructure. Applying this philosophy to AI seemed logical and desirable. Meta with Llama, Mistral with its Apache 2.0 models, and now DeepSeek with MIT seem to give form to that ideal.

But the open-source community finds itself facing a question it had not really anticipated: what to do when an open model is produced by an organization whose government is a strategic adversary of the democratic West? Freedom of code is a value — but is it absolute when that code can be used to reinforce an authoritarian system, train surveillance models, or create dependencies on infrastructure controlled by Beijing? The open-source community has no consensus answer to this question.

China's paradox: open models, closed market

There is a stinging irony in the situation: DeepSeek publishes its models under an MIT license — accessible to everyone in the world — while China blocks access to most Western AI models on its territory. GPT, Claude, Gemini are not accessible directly in China without a VPN. The Chinese AI market is protected by the Great Digital Wall. But DeepSeek can freely embed itself in the infrastructure of American, European, Japanese, and Korean companies.

This is a remarkable strategic asymmetry. China protects its domestic AI market while offering its own models to the rest of the world at a price that Western companies struggle to match. This is not free trade — it is a digital version of mercantilism, where strategic technologies are protected at home and massively exported abroad to create dependencies and erode competitive advantage.

Political responses: from outright bans to regulation

The American approach: between suspicion and use

In the United States, the response to DeepSeek has been ambivalent. On one hand, some members of Congress called for a security risk assessment related to the use of Chinese models in government agencies and critical infrastructure. The Department of Defense and several federal agencies have restrictive policies on the use of software developed by companies subject to foreign intelligence laws. On the other hand, Silicon Valley continues to integrate DeepSeek into its ecosystems, and many American developers use the model daily.

The most substantive response has been on the investment side: the United States is accelerating domestic open-source model initiatives — the DARPA trusted AI project, Meta's investments in Llama, and the efforts of several startups to develop open models comparable to DeepSeek V4 Pro under jurisdictions that carry no obligation to cooperate with a foreign government. The goal: provide the global community with a competitive open-weight alternative that does not carry the provenance risk of a Chinese company.

Europe: caught between technological dependence and regulation

Europe has a different response, shaped by its AI Act regulatory framework. Large general-purpose models — which DeepSeek V4 Pro would qualify as — are subject to transparency and compliance requirements. Companies deploying these models in high-risk applications must conduct conformity assessments. But European regulation does not distinguish between a model developed in San Francisco and one developed in Hangzhou — geopolitical provenance is not a criterion under the AI Act.

France with Mistral AI perhaps has the most coherent response: developing a world-class European open-source model that offers the same performance and accessibility advantages as DeepSeek, but under a jurisdiction that does not impose obligations to cooperate with a foreign government. Mistral Large 3 under Apache 2.0 is competitive but has not yet reached the size and performance of DeepSeek V4 Pro. Closing this gap should be a European technological sovereignty priority.

The real question: what is digital sovereignty?

Algorithmic dependence: the new strategic risk

The conversation about digital sovereignty has long focused on data — who holds European citizens' data, who has access to servers, who can monitor communications. DeepSeek V4 Pro adds a new dimension: algorithmic dependence. When a growing share of an economy's decision-making infrastructure — from recommendation systems to coding tools to legal and medical assistants — rests on models developed under an adversarial jurisdiction, a form of structural dependence sets in that goes beyond the data question.

This is not an argument to reject all Chinese models. It is an argument to develop credible alternatives. Diversifying AI providers — like diversifying energy suppliers after dependence on Russian gas — is a resilience strategy, not an ideological posture. The guiding principle should be: use DeepSeek where appropriate, but do not depend on it for critical functions, and actively invest in sovereign alternatives.

The missed opportunity of the major democracies

The real critique that can be leveled at Western governments in the face of DeepSeek V4 Pro is not that they "allowed" China to produce this model. It is that they did not invest early enough in competitive open-source models from their own ecosystems. European and American governments largely funded closed proprietary AI — OpenAI, Anthropic, Google DeepMind — while underinvesting in open, sovereign AI projects. DeepSeek fills a void that democracies could have occupied had they had the strategy and investment will.

The response to DeepSeek V4 Pro cannot be fear or outright bans. This model is already embedded in thousands of infrastructures worldwide. The credible alternative is investment: in world-class Western open-source models, in training engineers skilled in model architecture, in regulatory frameworks that distinguish between models based on their provenance and risk level. This is an ambitious political program. And it is the only one capable of effectively meeting the challenge posed by the Chinese open-source AI model.

DeepSeek and the AI competition: where do things stand?

DeepSeek V4 Pro: real performance and documented limits

DeepSeek V4 Pro sent shockwaves through the global artificial intelligence community at its release. Its performance on academic and coding benchmarks was comparable to — and on some tasks superior to — models like GPT-4o or Claude 3.5 Sonnet, at a fraction of the declared training cost. The demonstration raised a fundamental question: can world-class AI models be developed without the massive resources of Google, Meta, or OpenAI?

The honest answer is nuanced. On certain tasks — mathematical reasoning, coding, technical text comprehension — DeepSeek V4 Pro is genuinely competitive with the best American models. On others — low-resource language processing, complex multimodal tasks, safety and alignment — American models retain an advantage. And the training cost figures reported by DeepSeek remain contested by independent experts who estimate the real cost is likely higher than officially communicated.

The open-source ecosystem: an ambivalent blessing

DeepSeek's decision to publish its models as open-source — at least certain versions — multiplied their impact. Thousands of developers worldwide downloaded, fine-tuned, and deployed these models in their own applications. Within months, DeepSeek became a foundational component of the global AI ecosystem, embedded in applications ranging from coding to medicine to education.

This surge in open-source AI creates an unprecedented situation: world-class AI capabilities are now accessible to actors — states, businesses, individuals — who lacked the means to develop them independently. This is a democratization of AI with real positive effects. But it is also a proliferation of powerful capabilities without the safety guardrails that a responsible company would integrate — anti-disinformation filters, protections against malicious uses, alignment mechanisms.

AI regulation in Europe: between ambition and pragmatism

The European AI Act: first assessment of a pioneering regulation

The European AI Act, progressively entering into force since 2024, is the world's first legislation to regulate artificial intelligence in a systematic and binding manner. It establishes a risk-based classification — from unacceptable-risk systems (prohibited) to high-risk systems (subject to strict obligations) to limited-risk systems (subject to transparency obligations). This architecture is conceptually coherent and has inspired other legislators around the world.

But the first implementation assessments reveal predictable tensions. The scope of "high-risk systems" is defined broadly enough to encompass a significant proportion of commercial AI applications. The resulting documentation, audit, and compliance obligations represent an administrative burden that large companies can absorb but which may discourage SMEs and startups. The question of the competitiveness of the European AI ecosystem against American and Chinese players, who face fewer regulatory constraints, is real.

Balancing innovation and security: the central challenge of technology regulation

The trade-off between innovation and security is the central challenge of any technology regulation. If regulation is too strict, it stifles innovation and creates competitive disadvantages. If it is too lax, it enables abuses and harms that ultimately justify even more severe regulatory reactions. The optimal regulatory zone is narrow and must constantly adapt as technology evolves.

The European Union has chosen a more cautious approach than the United States — which has so far favored self-regulation and voluntary guidelines — and more constraining than China — which regulates AI primarily to maintain political control rather than protect citizens. This middle position has real merit in terms of fundamental rights protection. Its challenge is to demonstrate that it is compatible with building a competitive European AI ecosystem.

Artificial intelligence as a major geopolitical issue

The AI race: who is ahead and why it matters

The global competition in artificial intelligence is often framed as a binary US–China race. The reality is more nuanced. The United States retains an edge on frontier models — the most powerful systems like GPT-4o, Gemini Ultra, or upcoming models from OpenAI and Anthropic. China has closed part of the gap with DeepSeek and other domestic models, and holds specific advantages in surveillance and facial recognition AI. Europe maintains a presence in academic research and a few champions like Mistral AI, but remains structurally behind.

What matters geopolitically is not who has the most impressive model on a benchmark — it is who controls the infrastructure, training data, and critical applications. A country that depends on AI developed under an adversarial jurisdiction for its medical, judicial, military, or economic decisions carries a strategic vulnerability comparable to a country dependent on foreign energy controlled by a potential adversary.

Military applications of AI: a Rubicon crossed

Military applications of artificial intelligence represent the most sensitive dimension of this competition. AI systems for automatic target recognition, intelligence analysis, cyber defense, military logistics, and potentially decision support in combat operations — these applications are transforming the very nature of warfare. The war in Ukraine has been a laboratory of military AI innovation: autonomous drones, AI-assisted targeting systems, real-time mass analysis of satellite data.

Military powers that most effectively integrate AI into their armed forces will enjoy a growing operational advantage. This is a reality that American, European, and Chinese militaries have fully integrated into their planning. The question of the ethical and legal limits of autonomous weapons systems — which can select and engage targets without human intervention — is one of the most urgent and least covered debates in contemporary international security.

Conclusion: Neither demonization nor naivety

DeepSeek V4 Pro is not a singular threat — it is a symptom

DeepSeek V4 Pro under MIT license is neither a gift from Beijing with no ulterior motives, nor a militarized Trojan horse to be rejected outright. It is a remarkably high-performing AI model, developed by a private Chinese company within a legal and geopolitical framework that imposes specific constraints on sensitive uses. The appropriate distinction is not China/non-China — it is critical use/non-critical use, self-hosting/hosted API, public infrastructure/personal application.

This model is a symptom of a broader reality: China has developed AI capabilities competitive with the world's best, and it is deploying them with an open-diffusion strategy that maximizes global adoption and creates dependencies. The West must respond to this challenge with equal strategic intelligence — by investing in its own champions, by regulating risk-bearing uses in a targeted manner, and by refusing both the naivety that ignores risks and the paranoia that blocks innovation.

The imperative of active technological sovereignty

Technological sovereignty is not digital isolationism. It is the capacity to choose one's dependencies with full awareness. Faced with DeepSeek V4 Pro, the West has a choice between two postures: passive dependence on a powerful AI model developed under an adversarial jurisdiction, or active investment in sovereign alternatives that offer the same advantages without the same risks. This choice cannot be left to the market alone — it requires a deliberate political strategy. And the decisions made now will determine the dependencies of the global digital economy in ten years.

By Maxime Marquette, columnist

Columnist's transparency note

Position and biases

This editorial reflects my conviction that technological sovereignty is a strategic value that democracies must actively defend. I am not an expert in AI model architecture, and my assessment of security risks rests on public sources — security researcher reports, think tank analyses, and specialized media coverage. My anti-authoritarian bias regarding China is deliberate but does not target individual Chinese developers and researchers.

What I do not know

I cannot independently assess whether backdoors or surveillance mechanisms have been embedded in DeepSeek V4 Pro. Independent security analyses published to date have not found evidence of such mechanisms in the open-source model weights. This question remains open.

Sources

Primary sources

Hugging Face — DeepSeek-V4-Pro official README (MIT license) — April 24, 2026

CNBC — China's DeepSeek releases preview of long-awaited V4 model — April 24, 2026

Bloomberg — DeepSeek Unveils Flagship AI Model a Year After Breakthrough — April 24, 2026

Secondary sources

AI Blogs — DeepSeek V4-Pro lands at 1.6T total under MIT license — May 26, 2026

Reuters — DeepSeek unveils new AI model tailored for Huawei chips — April 24, 2026

AI News — China Blocks Western AI Models as Silicon Valley Embraces DeepSeek — June 13, 2026

Forbes — China's DeepSeek V4 And Qwen Reshape The Open-Source AI Race — April 28, 2026