Hong Kong

Chinese artificial intelligence startup DeepSeek on Friday released preview versions of its latest major AI model, V4, as competition between China and the United States in advanced AI technologies intensifies.

The launch had been eagerly awaited by users and industry watchers looking to compare the model with leading American rivals such as OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini.

DeepSeek said its open-source V4 family includes “Pro” and “Flash” variants, featuring major upgrades in knowledge, reasoning, and “agentic” abilities — meaning the capacity to independently complete complex tasks and workflows.

The V4 model succeeds DeepSeek’s V3 system, introduced in late 2024. However, it was the company’s reasoning-focused model DeepSeek R1 that drew global attention in January 2025 after claims it delivered strong performance at lower cost than comparable Western models.

According to the company, its “V4 Pro Max” version performs better than GPT-5.2 and Gemini 3.0-Pro on several reasoning benchmarks, while trailing slightly behind newer versions GPT-5.4 and Gemini 3.1-Pro.

DeepSeek also said the V4 Pro model outperforms Claude Sonnet 4.5 in autonomous task execution and approaches the level of Claude Opus 4.5 based on internal testing.

Technology analyst Lian Jye Su of Omdia said benchmark data suggests the new model could be highly competitive against US rivals.

Marina Zhang of University of Technology Sydney described the rollout as an important moment for China’s AI sector, especially as nations seek greater self-reliance in strategic technologies.

DeepSeek offers free web and mobile chatbot services and promotes its systems as open source, allowing developers to modify and build on its underlying technology. The company said both V4 variants support a 1 million-token context window, a sharp jump from V3’s 128,000-token limit.

Still, some experts urged caution. Ivan Su of Morningstar said V4 appears solid but may not be as disruptive as the earlier R1 launch, noting that stronger independent testing is needed.

READ MORE: Assam-born Fazlur Rahman finds cancer's link with environment

The update also comes amid continuing tensions over allegations from OpenAI and Anthropic that Chinese firms used “distillation” methods to learn from US models. China has rejected those claims and accused the US of unfairly targeting Chinese companies.