open weights

A model whose trained parameters you can download and run, without necessarily getting the training data, the code that made it, or a licence that lets you do what you like. The term arose because 'open source' was being used for models that were none of those things. Meta's LLaMA leaked in March 2023, DeepSeek released weights in 2025, and OpenAI, after years, published two in August 2025.

The Ampere-generation GPU announced in May 2020 that trained GPT-3 and most of the models that followed until the H100 arrived; the chip whose export to China was restricted in October 2022, and whose stockpiles in Chinese hands, reportedly including DeepSeek's, became a story in themselves. Retired from the cutting edge and still working everywhere else.

The proposition that the money going into AI, chips, data centres, power stations and start-ups with no revenue, exceeds anything the technology will return, and that the correction will be large. Nvidia passed a trillion dollars in 2023 and five trillion in 2025; the comparison to 1999 was made weekly from 2024. The word is used by people who are still buying.

The US rules, first issued in October 2022 and tightened in October 2023, that restricted the sale of advanced AI chips to China, producing a sequence of chips designed to sit just under the line (the A800, the H800, the H20), a smuggling economy, and in 2025 a policy reversal in which sales were permitted in exchange for a reported share of the revenue. The internet's entry for the chip as a diplomatic instrument.

The wave of government restrictions on the Chinese chatbot DeepSeek after its January 2025 release: Italy removed the app from stores, Australia and Taiwan banned it from official devices, South Korea suspended downloads, and US agencies including NASA and the Navy told staff not to use it. Several states followed. The internet cared because the same weeks had made DeepSeek the most downloaded app in the world.

A Monday in January when a Chinese model that cost a fraction of the American ones topped the app store and wiped hundreds of billions off chip stocks. Named like a Sputnik moment, memed like a stock crash: whale emojis everywhere.

The Chinese reasoning model released in January 2025 with open weights, a permissive licence and its chain of thought showing, which put its app at the top of the US App Store within a week and took a reported near six hundred billion dollars off Nvidia's market value in a single Monday. The record files the market event separately under DeepSeek Moment; this is the model.

The DeepSeek base model released on Boxing Day 2024 whose paper stated the final training run had cost around 5.6 million dollars in rented GPU time, a figure that was quoted for a month as the price of the whole company and reset the industry's sense of what a frontier model should cost. The record notes R1 was built on it four weeks later.