What happened
On August 5, 2025, OpenAI released the gpt-oss family — the company's first open weights since GPT-2 (2019): gpt-oss-120b (117B parameters) and gpt-oss-20b (21B), both MoE models with 4-bit MXFP4 quantization, under the Apache 2.0 license (OpenAI).
The headline: gpt-oss-120b is close to parity with o4-mini on the main reasoning benchmarks while running on a single 80GB GPU; gpt-oss-20b is comparable to o3-mini and runs on devices with 16GB of memory (TechCrunch). Simon Willison called the models "really good" after his first tests (simonwillison.net).
Why it matters
OpenAI kept its weights closed for six years — and the reversal came under pressure from Chinese open source (DeepSeek, Qwen, GLM). Apache 2.0 with no restrictions on commercial use raised the bar sharply for everyone: "openness" is a competitive field again, not a gesture.
What it means for vibe coding
Local coding agents became realistic on mid-range hardware: a 20b model on a laptop with 16GB is an autonomous assistant with no code leaving your machine. For sensitive projects, pairing "local gpt-oss + a cloud frontier model for the hard parts" turned into a working pattern.