Beijing: Chinese AI startup DeepSeek has launched its latest large language models, DeepSeek V4 Flash and DeepSeek V4 Pro, marking a major upgrade over its previous V3.2 and R1 systems that previously disrupted global AI markets.
The company said both are open-weight models featuring context windows of over 1 million tokens, enabling processing of extensive documents or full codebases. The V4 Pro model contains 1.6 trillion parameters (49 billion active), making it one of the largest open-weight AI models, surpassing competitors such as Moonshot AI’s Kimi K2.6 and MiniMax’s M1, and more than doubling its predecessor V3.2.
DeepSeek said the V4 Flash version offers a more efficient and cost-effective option with fewer parameters. The models are optimized for AI agent tools and compatible with Huawei chips, including Ascend SuperPoD systems.
The company highlighted that V4 reduces computing and memory costs significantly while improving long-context performance, a development experts describe as an industry “inflection point.” Analysts say it could expand advanced AI use beyond research labs into mainstream applications.
A preview version has been released, though no final rollout date has been announced. The launch comes amid intensified U.S.–China tensions over AI leadership, with both sides accusing each other of technology theft and competition in the rapidly evolving sector.
