DeepSeek V4 Review

DeepSeek V4 shipped on April 24, 2026 as an open weight mixture of experts family in two sizes, V4 Pro and V4 Flash, with the weights posted to Hugging Face under the MIT license. On coding it is the headline: V4 Pro scores 93.5 on LiveCodeBench and 80.6 percent on SWE-bench Verified, matching or beating the previous generation of closed frontier models, and it sits alongside them on math and question answering too.

The staggering part is price. The hosted API runs V4 Pro at about forty four cents in and eighty seven cents out per million tokens, and V4 Flash at fourteen and twenty eight cents, an order of magnitude below the Western flagships, and you can also run the weights yourself. The limits are the familiar ones for this class. It is a strong reasoner and coder but not quite the top of the agentic charts, the polish and safety tooling lag the big labs, and self hosting a large mixture of experts model is not trivial. There is no DeepSeek R2 despite the rumours.

This is the value pick of the year for developers who want to build on near frontier coding without a frontier bill, for anyone who needs on premise or private weights, and for high volume workloads where the price difference compounds. If you want a consumer grade assistant with a polished app and enterprise support, look to the closed models instead.