DeepSeek is reported to be preparing to unveil its latest large language model: V4. This release, anticipated within days, is set to be DeepSeek’s first major product launch in over a year.

V4 is expected to be a multimodal model, capable of generating and processing not only text but also images and video. The timing of the V4 debut is significant as China’s annual “Two Sessions” meetings (high-profile parliamentary gatherings that shape national policy) begin on 4 March.

DeepSeek has believed to have withheld pre-release access to its forthcoming V4 model from American chip giants Nvidia and AMD. This decision marks a sharp departure from the industry norm of collaborating with leading US chipmakers to optimise performance before launch.

Instead, DeepSeek has prioritised domestic hardware, granting early access exclusively to Chinese suppliers such as Huawei Technologies. This strategy has provided Chinese processors with a crucial advantage, allowing them several weeks to fine-tune the software and reinforce China’s drive to reduce dependence on US semiconductor technology.

American labs accuse Chinese rivals of illicit model replication

Last month we reported how Anthropic accused DeepSeek of conducting what it described as “industrial-scale campaigns” to extract the capabilities of its Claude model illicitly.

Anthropic said it had “identified industrial-scale campaigns by three AI laboratories, including DeepSeek, Moonshot, and MiniMax, to extract Claude’s capabilities to improve their own models illicitly”.

The company claims the three labs made over 16 million interactions with Claude using roughly 24,000 unauthorised accounts, breaching its terms and access rules.

Distillation involves training a smaller AI model using the outputs of a more advanced one. Anthropic recognises it as standard practice for developing efficient versions of existing models, but cautions it can be exploited to quickly and cheaply replicate competitors’ capabilities. The company describes the issue as urgent and global, warning that these activities are rapidly advancing and require swift, coordinated action from industry, policymakers, and the international AI community.

The claims follow similar allegations from US rival OpenAI, which earlier this month told the US House Select Committee on China that DeepSeek may have illegally distilled ChatGPT models over the past year.

OpenAI alleges that DeepSeek’s competing model used “distillation” to copy its technology. Tighter security measures now restrict access to sensitive algorithms and products. OpenAI also said the company has increased security at data centres and hired more cybersecurity staff.

Anthropic said it had attributed each campaign to a specific company with “high confidence” using IP address correlation, request metadata, infrastructure indicators, and verification from industry partners.

Anthropic stated that each campaign targeted Claude’s advanced features (agentic reasoning, tool use, and coding), demonstrating intentional extraction of capabilities rather than typical use. DeepSeek’s campaign involved over 150,000 interactions, focusing on reasoning, rubric-based reinforcement learning, and creating “censorship-safe” responses.

RELATED STORIES

Anthropic blasts China’s top AI labs for stealing its model

DeepSeek in spotlight: Nvidia’s China AI support draws fresh US scrutiny