DeepSeek V4 Set to Launch This Week as Anthropic Accuses It of Industrial-Scale Model Theft
DeepSeek's first major AI release in over a year arrives amid Anthropic allegations that the Chinese lab used 24,000 fake accounts and 16 million queries to extract Claude's capabilities.
Overview
DeepSeek, the Hangzhou-based AI lab whose V3 model rattled Wall Street in January 2025, is expected to release its next major model as early as this week. The timing is politically charged: the launch of V4 is reportedly coordinated to precede China’s annual parliamentary “Two Sessions” meetings beginning March 4, positioning the model as a showcase of domestic AI capability. It arrives alongside a formal accusation from Anthropic that DeepSeek and two other Chinese AI companies engaged in what the company calls “industrial-scale” extraction of its proprietary Claude model’s capabilities—a legal and technical dispute that could define how the global AI industry regulates competitive intelligence gathering.
What We Know About DeepSeek V4
According to reporting by the Financial Times and CNBC, DeepSeek plans to release V4 this week, marking the company’s first major model launch since the R1 reasoning model debuted in January 2025. Sources familiar with the matter told CNBC that V4 is a “multimodal” system capable of generating text, images, and video—a significant architectural expansion beyond the text-focused V3 series.
The model is reportedly built as a trillion-parameter Mixture-of-Experts architecture with approximately 32 billion active parameters, a 1 million-token context window, and native multimodal processing integrated during pre-training rather than added as a post-hoc module. Unlike previous DeepSeek releases, V4 was optimized for Chinese-made semiconductors: sources told the Financial Times that the company worked with Huawei and Cambricon, two domestic chipmakers, to tune V4 for their latest hardware—and withheld early access from Nvidia and AMD.
The hardware optimization carries strategic weight. DeepSeek’s V3 model was notable in part for its reported training cost of under $6 million—far below comparable U.S. efforts—achieved partly through clever use of lower-powered Nvidia chips. V4’s apparent pivot to Huawei Ascend and Cambricon processors suggests the lab is stress-testing whether frontier-class models can be developed and deployed without U.S.-origin silicon, a capability that would materially blunt the effect of American export controls if demonstrated at scale.
CNBC reported that internal benchmarks suggest V4 may outperform Claude and ChatGPT on long-context coding tasks, though DeepSeek has not published independent evaluations or a technical report ahead of the release.
The Anthropic Distillation Accusation
The V4 launch is shadowed by a public accusation that throws into question how V4’s capabilities were built.
On February 23, Anthropic published a blog post alleging that DeepSeek, along with Chinese AI developers Moonshot AI and MiniMax, had conducted “industrial-scale campaigns” to extract capabilities from its Claude model through a process known as distillation—a technique in which a smaller or less capable model is trained on the outputs of a more powerful one rather than on raw data. As TechCrunch reported, the accusations were disclosed exclusively, timed to coincide with an active U.S. government debate over AI chip export controls.
According to Anthropic, the three companies collectively generated over 16 million exchanges with Claude using approximately 24,000 fraudulent accounts—accounts that Anthropic said were used to bypass geographic restrictions and terms-of-service enforcement. The company said it has since “cut off known access points” but stopped short of filing lawsuits.
The breakdown, as reported by Fortune and CNBC, attributes more than 13 million exchanges to MiniMax, over 3.4 million to Moonshot AI, and a smaller share to DeepSeek. However, Anthropic noted that DeepSeek’s queries were specifically targeted: the company focused on reasoning tasks, rubric-based grading useful for reinforcement learning reward models, and rewrites of politically sensitive material—suggesting a deliberate effort to extract capabilities relevant to training advanced reasoning and alignment systems, not simply to use Claude for ordinary tasks.
Bloomberg reported that Anthropic is urging “rapid, coordinated action” among AI companies and policymakers, framing the distillation campaigns as a national security concern given the potential for foreign state actors to acquire frontier AI capabilities without developing them from scratch.
None of the three accused companies has publicly responded to the specific allegations.
What We Don’t Know
Several critical uncertainties surround both the V4 release and the distillation controversy.
On the model side, DeepSeek has not released benchmark results, a technical report, or an official release date. The timeline remains based on anonymous sourcing, and previous windows for the release—reportedly including mid-February and late February—passed without a launch. The actual capability gap between V4 and current frontier models from Anthropic, OpenAI, and Google will not be verifiable until independent evaluations are published.
On the legal and policy side, the status of distillation under intellectual property law is unresolved. Anthropic has not filed lawsuits, and the legal theory underpinning a distillation claim—whether training on another model’s outputs constitutes infringement—has not been tested in court. Critics have also noted the irony of Anthropic’s complaint, given that the company recently settled a $1.5 billion copyright dispute with authors over its own training data practices.
Whether the alleged distillation of Claude materially contributed to any specific DeepSeek model’s capabilities is also unknown. Distillation can meaningfully transfer specific skills—particularly in reasoning and coding—but the degree to which 16 million exchanges influenced a trillion-parameter system remains an empirical question that neither Anthropic nor independent researchers have answered publicly.
Analysis
The confluence of these two events—a major Chinese AI model release and a public accusation of systematic capability extraction—reflects a structural tension in the global AI industry that is likely to intensify.
The case for the significance of DeepSeek V4 rests not only on its technical specifications but on the geopolitical signal embedded in its hardware choices. If a frontier-class multimodal model can be trained and deployed on Huawei and Cambricon silicon, the effectiveness of U.S. chip export controls as an AI containment strategy is substantially reduced. The previous DeepSeek shock—the V3 release in January 2025—demonstrated that the cost and hardware gap between Chinese and American frontier models was smaller than assumed. V4 would test whether that gap has closed further, and whether it can be maintained even on domestically produced chips operating under export restrictions.
Anthropics’ distillation complaint, meanwhile, reveals a gap in the current regulatory and legal framework. The industry has no established norms around what constitutes impermissible competitive intelligence gathering from AI systems, and the techniques at issue—synthetic data generation, reinforcement learning from AI feedback—are standard tools used by essentially every major lab. The accusation implicitly asks regulators and courts to distinguish between legitimate research use and systematic capability extraction, a line that is technically and legally difficult to draw.
The timing of Anthropic’s disclosure—released publicly alongside active U.S. export control debates—suggests the company sees the distillation campaigns as leverage for tighter restrictions on AI service access as well as chip exports. Whether policymakers respond will depend heavily on how verifiable the claims are and whether other U.S. labs have experienced similar targeting.
For observers tracking the competitive AI landscape, the more immediate question is whether V4’s release, when it arrives, will produce the same market disruption as V3 did in 2025. The context in 2026 is different: U.S. tech companies have already recalibrated their cost assumptions following the first DeepSeek shock, and the $650 billion in AI infrastructure spending projected for 2026 was made with awareness that capable Chinese models exist. Whether V4 demonstrates a step-change sufficient to shift those assumptions again remains to be seen.