DeepSeek-V3
A Chinese laboratory opened a 671-billion-parameter model of which 37 billion work on each token, and named its training cost: about 5.6 million dollars.
Why it matters
The frontier was reached for the first time on a budget smaller than a mid-sized laboratory's annual spending, by a company under export restrictions.
The report described the engineering frankly: mixed-precision arithmetic, a custom protocol for exchange between accelerators, training on chips cut down for the Chinese market. The stated cost covered only the final run and left out research, but even allowing for that the distance from the figures American laboratories quoted was striking. The weights were published under a licence permitting commercial use. A month later the same laboratory's next model would knock down the market value of American chipmakers.