DeepSeek just previewed two new models that the company says outperform its previous V3.2 release on both efficiency and reasoning benchmarks. The Chinese AI lab published details on April 24, positioning the new models as its closest match yet to top closed-source systems like GPT-4o.
Both models benefit from architectural improvements that make them faster and cheaper to run than V3.2 while scoring better on reasoning benchmarks - tests that measure how well a model works through multi-step logic problems, like math proofs or code debugging. DeepSeek says the new models have "almost closed the gap" with current frontier models, open-source and proprietary alike.
That's a significant claim. DeepSeek V3 and its reasoning model R1 already surprised the industry earlier this year by matching ChatGPT-class performance at substantially lower cost. If the preview models deliver on these benchmark results in real-world use, it extends a now-familiar pattern: a Chinese lab building competitive models faster and cheaper than most expected. DeepSeek has not announced a release date for the full versions yet.