Inside DeepSeek V4's Reasoning Stack

DeepSeek V4 is the clearest signal yet that reasoning quality and inference cost are no longer locked in a fixed tradeoff. Here is what the public documentation tells us about how they did it.

#The architecture: sparse, again

Like several of this generation's flagships, V4 is a mixture-of-experts model. Only a slice of its parameters activate per token, which keeps per-token compute low even as total capacity grows.

#Reasoning as a trained behavior

The headline results come from post-training, not pre-training. The documented recipe has three stages:

1. Cold-start reasoning traces — curated long-form chain-of-thought data
2. Rule-based reinforcement learning — rewards for verifiable answers in math and code, no human labelers required
3. Distillation back into the base policy — so fast inference keeps most of the reasoning gains

#Why verifiable domains first

Math and code have automatic graders: an answer is right or wrong, a test passes or fails. That makes reinforcement learning cheap and scalable. Open-ended tasks still rely on the older supervised pipeline.

#The cost angle

V4's per-token price sits far below closed frontier models, and the sparse architecture means self-hosting is realistic for organizations with moderate GPU budgets. Combined with open weights, that changes who can deploy frontier-class reasoning.

#What it means

The moat is moving from model access to training pipelines. If reasoning can be taught with verifiable rewards, the labs with the best reward infrastructure — not the biggest models — set the pace.

Full registry entry on the model radar.