DeepSeek is releasing its V4 multimodal AI model this week, timed to coincide with China's annual Two Sessions parliamentary meetings starting March 4. The trillion-parameter model represents a major leap — capable of generating text, images, and video natively — and marks the Chinese startup's first major launch since the R1 reasoning model shook Silicon Valley in January 2025.

In a striking strategic move, DeepSeek deliberately excluded Nvidia and AMD from V4's optimization pipeline, instead building its inference stack around Huawei Ascend and Cambricon chips. With approximately 32 billion active parameters per token and a 1-million-token context window, V4 reportedly matches frontier Western models on coding benchmarks while running entirely on Chinese-made hardware.

Leaked benchmarks suggest V4 scores around 90% on HumanEval and above 80% on SWE-bench Verified, putting it in the same tier as Claude Opus 4.6 and GPT-5.3 Codex. If confirmed, the model would demonstrate China can produce frontier AI without relying on Western chip supplies — a milestone with significant geopolitical implications.