Chinese AI startup DeepSeek is back with its most anticipated release in over a year. DeepSeek V4 marks the company’s most significant release since R1, the reasoning model that stunned the global AI industry in January 2025 with its strong performance and efficiency, transforming DeepSeek from a little-known research team into China’s best-known AI company almost overnight.
Two Models, One Million Tokens
The V4 series comes in two flavors: DeepSeek-V4-Pro, a massive model with 1.6 trillion total parameters (49 billion activated), and DeepSeek-V4-Flash, a leaner variant with 284 billion parameters (13 billion activated). Both support a context length of one million tokens.
V4-Pro is built for coding and complex agentic tasks, while V4-Flash is designed to be faster and cheaper to run. Both versions offer reasoning modes, where the model works through problems step by step.
Architectural Leap
Under the hood, V4 introduces significant engineering advances. The V4 family builds on DeepSeek’s Mixture-of-Experts architecture with a new hybrid attention mechanism, achieving a 73% reduction in per-token inference compute and a 90% reduction in memory burden compared to its predecessor, DeepSeek-V3.2.
Capable — and Cheap
DeepSeek claims V4 achieves world-class reasoning, leads all open models in math, STEM, and coding, and delivers the best agentic coding capability among open-source models. DeepSeek That said, the company acknowledges its model still trails the very top closed-source systems like Gemini.
The pricing, however, is hard to ignore. DeepSeek-V4-Pro is priced at $1.74 per million input tokens — roughly one-seventh the cost of GPT-5.5 and about one-sixth the cost of Claude Opus 4.7.
Running on Chinese Chips
One of V4’s more geopolitically significant details is its hardware. DeepSeek partnered with Chinese tech giant Huawei, using its “Ascend 950” chips to power V4 — a notable shift from R1, which relied on Nvidia hardware. Analysts say this could accelerate AI adoption domestically and reduce dependence on US chip suppliers.