OpenAI’s Surprise Move: A New Open-Source Model You Can Run on Your Laptop

OpenAI has unveiled GPT-OSS — its first open-weight model release in six years — a model that can be freely downloaded, executed on a laptop, and customized for individual use. The model is available in two variants: a full-scale version with 120 billion parameters and a lighter 20-billion-parameter version. Both are distributed under the Apache 2.0 license and are accessible via Hugging Face, Databricks, Azure, and AWS.

The larger model is designed to operate on a single Nvidia GPU and delivers performance comparable to o4-mini, while the smaller version — aligned with o3-mini — can run on systems with just 16GB of video memory. According to the company, both models are already capable of reasoning tasks, code generation, agent orchestration via the OpenAI API, and even internet access.

This release marks OpenAI’s first step back into the open-source space after a lengthy hiatus. Company CEO Sam Altman had previously cited security concerns as the primary reason for withholding such models. However, following the rapid rise of open-source initiatives like DeepSeek, he conceded that OpenAI had found itself “on the wrong side of history.”

Now, OpenAI aims to regain the attention of developers who have increasingly gravitated toward open models due to their flexibility and low cost. OpenAI researcher Chris Kuo noted during a briefing that “most OpenAI customers are already using open models,” and GPT-OSS is intended to help bridge that divide.

The company emphasizes that GPT-OSS has undergone “the most rigorous testing” of any OpenAI model to date, including assessments by independent auditors to mitigate potential risks related to cybersecurity and bioweapon misuse. The model incorporates a reasoning trace mechanism that allows observers to follow its logical steps and identify distortions or manipulations. Nevertheless, like all OpenAI models, GPT-OSS does not disclose its training data and only produces textual outputs.

While OpenAI has not released direct benchmark comparisons against rivals such as Llama, DeepSeek, or Google’s Gemma, the company asserts that GPT-OSS performs competitively in coding tasks and on exams like Humanity’s Last Exam, holding its own against proprietary counterparts.

Although OpenAI has not committed to regular updates for GPT-OSS, it hopes the model will be embraced by independent developers and startups that prioritize transparency and data sovereignty. “When you lower the barrier to entry, innovation flourishes. People begin to experiment and create the truly unexpected,” remarked OpenAI co-founder Greg Brockman.