OpenAI Releases First Open-Source Language Models Since GPT-2

Unveiled yesterday, OpenAI's gpt-oss-120b and gpt-oss-20b mark the company's return to open-source AI after six years. Unlike typical language models that simply generate text, these reasoning-focused models can think step-by-step, use tools like web search and code execution, and run entirely on local hardware - no cloud required.

Building on OpenAI's breakthrough reasoning architecture from the o-series, both models feature mixture-of-experts design for maximum efficiency. The 120B version runs on a single 80GB GPU while the smaller 20B needs just 16GB RAM. Released under Apache 2.0 license, they're designed for developers who want full control over their AI infrastructure.

What's driving the excitement? Complete transparency and ownership. Your own reasoning AI, locally hosted, infinitely customizable. The tech democratizes advanced AI capabilities, from startups building specialized agents to governments requiring data sovereignty. Early benchmarks show gpt-oss-120b matching o4-mini performance on mathematical reasoning (96.6% on AIME 2024) and coding challenges, while the 20B version punches above its weight class.

Of course, challenges remain, both models are text-only, knowledge cutoff at June 2024, and prone to hallucinations compared to larger siblings. But OpenAI's new "worst-case fine-tuning" safety protocol passed external review, proving the models resist misuse even after adversarial training.

GPT-oss signals a future where cutting-edge AI isn't locked behind APIs, it's yours to build with.

Source: OpenAI