HONG KONG – Chinese artificial intelligence startup DeepSeek has launched preview versions of its latest flagship models, the V4 Flash and V4 Pro series, signaling a significant escalation in the global AI race. The company, which previously disrupted markets with its cost-effective AI systems, is now positioning its new offerings as the most powerful open-source platforms available, directly challenging established leaders such as OpenAI, Anthropic, and Google.
The V4 series introduces substantial architecture upgrades and optimization improvements, with a particular emphasis on advanced reasoning and agentic capabilities. A key innovation highlighted by DeepSeek is its Hybrid Attention Architecture, designed to enhance the AI's ability to retain context over long interactions. Furthermore, the models support a one million-token context window, allowing for the processing of entire codebases or extensive documents within a single prompt.
DeepSeek's latest release arrives a year after the company's R1 model caused significant market fluctuations by rivaling top-tier AI performance at a fraction of the cost. The new V4 models are claimed to surpass OpenAI's GPT-5.2 and Google's Gemini 3.0-Pro in standard reasoning benchmarks, while approaching the performance of GPT-5.4 and Gemini 3.1-Pro. In terms of agentic functions, the V4 'pro' version is said to outperform Anthropic's Claude Sonnet 4.5 and approach Opus 4.5.
The release has already seen Chinese chipmakers rally, with investors anticipating increased demand for local semiconductor products. However, DeepSeek has noted that the service capacity for the V4 Pro series is currently limited due to a computing crunch. The company anticipates a significant reduction in pricing once computing clusters powered by Huawei Technologies' Ascend 950 chips become operational in the latter half of 2026. DeepSeek is reportedly in discussions for its initial funding round with major tech firms Tencent Holdings and Alibaba Group Holding.
The competitive landscape is further intensified by ongoing accusations from competitors like Anthropic and OpenAI, who have alleged that DeepSeek has illicitly extracted capabilities from their models. In parallel, the U.S. administration has vowed to crack down on foreign entities, particularly those based in China, for exploiting U.S.-developed AI models. Despite these geopolitical tensions, DeepSeek's open-source approach aims to empower developers to modify and build upon its core technology, offering a distinct alternative to the proprietary models of its global rivals.