DeepSeek V4

DeepSeek released V4 on April 24, 2026, in two variants: V4-Pro, a 1.6 trillion parameter model with a one million token context window by default, and V4-Flash, a smaller and faster sibling. On most agentic benchmarks V4-Pro matched the leading closed models of the moment at roughly ten to thirteen times lower cost per output token. DeepSeek's own claim was that the model trailed the state of the art by three to six months while costing a fraction of the price.

The efficiency came from architecture rather than scale. V4 introduced Manifold-Constrained Hyper-Connections to hold training stable at extreme size, a hybrid attention system combining Compressed Sparse Attention and Heavily Compressed Attention that cut inference FLOPs by as much as 73%, and an Engram Conditional Memory module providing constant-time knowledge retrieval rather than paying attention cost over the full context.

This was the second time DeepSeek had produced this effect. R1, in early 2025, had broken the assumption that frontier capability required frontier spending on training. V4 broke the corresponding assumption about serving, showing that the cost of running a near-frontier model in production could be reduced by an order of magnitude without a matching loss in capability.

V4 changed what a closed frontier model was actually selling. When an open-weight system reaches within a few months of the state of the art at a tenth of the operating cost, the premium for the closed model is no longer capability in the abstract. It is the remaining months of lead time, the reliability of the surrounding platform, and the guarantees a vendor is willing to make. Every organization running significant inference had to re-examine whether it was paying for capability or for convenience.

The strategic consequence extended beyond pricing. A near-frontier model with published weights can be run inside a private network, inspected, fine-tuned, and audited in ways an API cannot. For regulated industries and for governments, that difference is often decisive regardless of benchmark scores. V4 arrived in the same year that European transparency obligations took effect, and the combination pushed serious infrastructure work toward models that could be held and examined rather than only called.