When DeepSeek released DeepSeek-R1 in January 2025, it blew everyone’s mind. R1 came surprisingly close to OpenAI’s o1 on several reasoning, math, and coding benchmarks, and DeepSeek released the weights under an MIT license along with a paper explaining much of its RL process. V3 had already shown that a Chinese lab could build a very capable model efficiently. R1 went further, showing that even advanced reasoning performance was no longer exclusive to a handful of frontier labs. They still had enormous advantages in compute, distribution, talent, and proprietary data. But the model itself was starting to look like a less reliable moat.
Now, DeepSeek Harness may be doing something similar one layer higher.
