DeepSeek Tests V4 Coding AI, Raising Stakes in U.S.-China Tech Rivalry

The Chinese artificial intelligence startup that jolted Silicon Valley with a cut-price reasoning model last year is preparing a new system aimed squarely at one of the tech industry’s most sensitive assets: its source code.

DeepSeek, a Hangzhou-based company backed by hedge fund High-Flyer, has begun internal testing of a next-generation large language model known as V4 that is designed for computer programming and software development, according to a report from The Information cited by Reuters on Jan. 9. The model is expected to launch publicly in mid-February.

People familiar with the product told The Information that V4 is optimized for writing and analyzing code, can process extremely long programming prompts and, in internal testing, has in some cases outperformed models from OpenAI and Anthropic on coding tasks. Reuters said it could not independently verify those claims and that DeepSeek did not respond to a request for comment.

If V4 performs as advertised, it would intensify competition in the fast-growing market for AI coding assistants, where tools such as GitHub Copilot, OpenAI’s GPT-4.1, Anthropic’s Claude and Google’s Gemini are already reshaping how software is written. It would also deepen scrutiny of DeepSeek, which U.S. officials have accused of helping China’s military and intelligence services and of trying to skirt American export controls on advanced chips—allegations the company has not publicly addressed in detail.

The emerging clash over V4 illustrates how AI models are moving beyond general-purpose chatbots into the infrastructure of software development itself, turning everyday developer tools into a new front in the technology and security rivalry between Washington and Beijing.

A coding model in a crowded field

V4 is described in The Information’s account as “focused on coding” with “strong coding capabilities” and breakthroughs in handling very long code inputs—the kind of sprawling, multi-file repositories common in large enterprise systems. That positions it directly against high-end systems from American and European developers that have made long context windows and repository-scale reasoning a key selling point.

OpenAI has promoted its GPT‑4.1 family, introduced in 2025, as offering substantial gains in coding performance over earlier models and the ability to process up to a million tokens of text—enough to ingest large codebases in one query. Anthropic’s Claude 3.5 Sonnet, released in 2024, features a 200,000-token context window and has scored strongly on standard programming benchmarks such as HumanEval. Google’s Gemini 2.5 Pro, rolled out with a similar focus on developers, supports million-token prompts and powers tools for writing, testing and deploying code.

Those systems are increasingly embedded directly into integrated development environments, terminals and continuous integration pipelines, where they can propose functions, refactor legacy software and generate tests. By aiming V4 at the same use cases, DeepSeek is signaling that it intends to compete not just on price but on the core capabilities that control how code is created and maintained.

So far, no technical paper or public benchmark has been released for V4, and neither The Information nor Reuters reported specific parameter counts or context-window sizes. The only performance figures made public are DeepSeek’s own internal test results, relayed secondhand. That leaves outside researchers and corporate security teams waiting to see whether the system can match Western competitors on widely used evaluations such as SWE‑Bench and HumanEval, and in real-world software projects.

From trading firm to AI price disrupter

DeepSeek was founded in July 2023 as an offshoot of High-Flyer, a quantitative hedge fund that had invested heavily in GPU clusters to support algorithmic trading. Company filings and public statements indicate that founder and chief executive Liang Wenfeng, who also leads High-Flyer, held a controlling stake in DeepSeek through holding companies.

Before Washington tightened export controls on advanced chips, High-Flyer and DeepSeek reportedly acquired around 10,000 Nvidia A100 graphics processors. As U.S. rules later restricted access to more powerful H100-class chips, the company said it shifted to training models on Nvidia’s China-compliant H800 processors and, in some cases, on domestic hardware.

DeepSeek gained global attention in early 2025 with DeepSeek‑R1, a “reasoning” model that it claimed could handle complex analytical tasks at a fraction of the cost of U.S. rivals. The company said usage of R1 was 20 to 50 times cheaper than OpenAI’s o1 line for comparable workloads. The release briefly sent shares of major Western tech firms lower as investors reassessed the economics of large language models.

Another flagship model, DeepSeek‑V3, received a major upgrade in March 2025 that improved performance on reasoning and coding benchmarks. An AI assistant app powered by V3 rose to the top of the free charts in Apple’s U.S. App Store that month, according to app-store rankings cited in media reports.

DeepSeek has also differentiated itself by publishing open-source variants of several models, allowing developers around the world to run them on their own hardware. That approach has attracted enthusiasm from researchers and startups but drawn concern from some regulators in China, where authorities require AI systems to comply with content rules emphasizing what officials call “core socialist values.”

Allegations of military links and data risks

As DeepSeek’s models have spread, U.S. officials have sharpened their criticism of the company.

In June 2025, a senior State Department official told Reuters that Washington believed DeepSeek had “willingly provided and will likely continue to provide support to China’s military and intelligence operations.” The official said the company appeared in more than 150 procurement documents tied to the People’s Liberation Army and China’s defense-industrial base, including contracts for AI-related services to military research institutes. Those records could not all be independently verified.

The same official said U.S. agencies were examining whether DeepSeek or related entities had tried to obtain Nvidia’s H100 chips—which are restricted for sale to China—and remote access to overseas data centers using shell companies in Southeast Asia. Nvidia has said its own review suggested DeepSeek trained its models on lawfully acquired H800 processors and that the company does not support any efforts to circumvent export rules.

American officials and some cybersecurity experts have also questioned how DeepSeek handles user data. Reports have linked parts of its backend infrastructure to China Mobile, a state-controlled telecoms carrier, and raised concerns that data from overseas users might be stored or processed in mainland China. DeepSeek has in several cases declined to answer detailed questions about its data practices.

DeepSeek has not been added to the U.S. Commerce Department’s Entity List, which would impose sweeping trade restrictions, but lawmakers have urged the Biden administration to consider tightening controls on AI firms alleged to support the Chinese military.

Censorship, bias and code quality

V4’s focus on programming comes against a backdrop of questions about how DeepSeek models behave when asked to write code for politically or ethically sensitive clients.

In September 2025, The Washington Post reported on tests carried out by cybersecurity firm CrowdStrike, which examined how one DeepSeek model responded to requests for industrial control systems software under different assumed end users. The researchers found that roughly 22.8% of the generated code samples were unsafe overall, but that the share rose to 42.1% when the hypothetical customer was described as the Islamic State extremist group.

The model was also more likely to refuse requests entirely when the prompts mentioned groups that Beijing regards as politically sensitive, including Falun Gong and Taiwan’s government, than for neutral or U.S.-based organizations. Experts quoted in the report said that pattern could reflect explicit alignment with Chinese content rules or unintentional bias inherited from training data drawn from a censored domestic internet.

In a separate development, Chinese telecom equipment maker Huawei announced in 2025 that it had co-developed a DeepSeek-based model trained on 1,000 of its Ascend chips that was tuned to block almost all content on topics deemed politically sensitive by Chinese regulators. Huawei described the project as a demonstration of how open-source AI models could be adapted to comply with national security and content regulations.

Those findings highlight a central question for V4: whether a coding engine developed under China’s regulatory and political environment could treat users, topics or regions differently in ways that affect not just text responses but the security and quality of software itself.

A new fault line in developer tools

The prospect of a powerful, low-cost coding model from DeepSeek underscores a dilemma for software teams around the world.

On one side, V4 could offer significant productivity and cost advantages if it delivers on claims of strong coding performance and long-context understanding comparable with or better than U.S.-built systems. For startups, independent developers and cash-constrained companies—particularly outside the United States and Europe—the economics could be compelling.

On the other side, corporate security officers and regulators are likely to scrutinize the idea of funneling proprietary code and architectural details through a model developed by a company that U.S. officials accuse of supporting a foreign military. Firms that work in defense, critical infrastructure, finance or other heavily regulated sectors may face legal or contractual limits on using tools that could expose source code or system designs to overseas providers.

The result could be a more fragmented global landscape for developer tools, with companies in the United States, Europe and allied countries gravitating toward domestic or closely vetted models, while organizations in China, parts of Asia and the Global South adopt DeepSeek and other Chinese systems.

For now, much about V4 remains unknown. DeepSeek has not publicly confirmed the model’s existence, timing or capabilities beyond what employees have shared anonymously with technology media. No independent evaluations have been released, and no regulator has issued guidance specific to the product.

Yet the direction of travel is clear. As AI systems like V4 move from answering questions to quietly writing and restructuring the code that runs economies and governments, decisions about which model to use are no longer just technical choices. They are decisions about where data flows, whose rules apply and which country’s technology stack an organization is willing to build on.