gpt-oss
OpenAI released two open-weight models under an Apache licence, its first since GPT-2; the larger runs on a single accelerator, the smaller on a device with 16 gigabytes of memory.
Why it matters
A company whose name had been a reproach about closedness for six years returned to open weights, under pressure from others releasing them.
GPT-2 in 2019 was at first withheld on safety grounds, and that set the tone for years. Here the 120-billion model reaches o4-mini's level on reasoning and fits on one 80-gigabyte accelerator, while the 20-billion model runs on a laptop. The Apache 2.0 licence carries no restrictions. The timing speaks: the decision matured over the six months in which R1 and Qwen3 made open weights a Chinese advantage. How many parameters. The names 120b and 20b are roundings: Hugging Face, which hosts the weights, counts 117 billion and 21 billion parameters in total, 5.1 and 3.6 billion active per token. The single 80-gigabyte accelerator for the larger model, 16 gigabytes for the smaller and the Apache 2.0 licence agree with the announcement in this source.