OpenAI Returns to Open Source After Six Years: gpt-oss Ships Under Apache 2.0

OpenAI released gpt-oss-120b and gpt-oss-20b on August 5, 2025 — two open-weight reasoning models under Apache 2.0, the company's first open weights since GPT-2 in 2019.

Per OpenAI, the 120b model runs on a single 80GB GPU while the 20b runs on 16GB edge devices; both are optimized for reasoning and agentic tasks, with weights on Hugging Face. Sam Altman called them 'the best and most usable open models in the world.'

Why the Turn After Six Closed Years

From GPT-2 (2019) to gpt-oss (2025), OpenAI built its commercial rise on six closed years — and turned the 'Open' in its name into an industry punchline. The immediate trigger was the R1 shockwave (see our coverage): once a frontier reasoning model shipped under MIT license for anyone to deploy, the safety-first case for closing weights lost its footing. Altman had conceded early in the year that OpenAI was 'on the wrong side of history' on open source — gpt-oss is that concession made concrete.

The Two-Track Calculus

The flagship stays closed and paid (GPT-5 shipped the same week — see our coverage); the open track gets reasoning-optimized small and mid models. That wins back developer mindshare and the self-hosting market without touching the API revenue base. The sizes are equally deliberate: a single 80GB GPU and 16GB edge devices sit exactly on the enterprise self-host and local-deployment demand gaps (see our local LLM deployment review).

The New Open-Source Balance

After gpt-oss, the open-weight race settled into a three-way tug: the U.S. camp with gpt-oss and Llama; Europe's Mistral; and China's DeepSeek, Qwen and Kimi setting the pace with a more aggressive release cadence (see our coverage). The debate moved from 'does open source matter' to 'whose open source is stronger' — and weight openness became a matter of national AI industrial policy.

Our Take

The real signal of gpt-oss is not OpenAI's change of heart but that open source has shifted from an ideological fight to a gateway fight: open models are the developer-ecosystem funnel and the ticket into enterprise self-hosted scenarios. For developers it is pure upside — as frontier labs compete on open weights, the waterline of locally deployable intelligence only rises.