DeepSeek-V4

Mon, 01 Jan 0001 00:00:00 +0000

DeepSeek-V4 is DeepSeek’s current generation, announced 24 April 2026. It ships as two models — deepseek-v4-pro at 1.6 trillion total parameters with 49 billion active, and a Flash line several times smaller — both with a million-token context, both under the MIT licence.

The generation matters for two reasons beyond its size. It is where DeepSeek finished merging its reasoning line back into its general line, so that thinking is a per-request setting rather than a separate model. And it is where the architecture that made the lab’s name, multi-head latent attention, stopped being what it ships.

DeepSeek-R1

Mon, 01 Jan 0001 00:00:00 +0000

DeepSeek-R1 is a reasoning model released by DeepSeek in January 2025, built on DeepSeek-V3-Base: 671 billion total parameters, 37 billion active per token. Its weights are published under the MIT licence.

It earns a page of its own for one result. Its predecessor DeepSeek-R1-Zero was trained to reason by reinforcement learning alone — starting from a base model that had never been shown a single worked example of step-by-step reasoning, with no supervised fine-tuning stage at all. Long chains of thought, self-correction and the habit of checking an answer before committing to it were not demonstrated to the model. They appeared during training, because they earned reward.

DeepSeek on Bitsy Wiki

DeepSeek-V4

DeepSeek-R1