OpenAI releases gpt-oss-120B and gpt-oss-20B under Apache-2.0-style license
For years the AI world has been tugging at a knot: big companies push ever-larger models, yet they usually hide the weights, saying safety and competition are at stake. Regulators have been nudging for more openness, arguing that public access might keep monopolies in check and let independent eyes examine the tech. In that setting OpenAI dropped two mixture-of-experts models, one around 120 billion parameters, the other about 20 billion, under a license that looks a lot like Apache 2.0.
It feels like a throwback to the GPT-2 moment when the organization first made its weights public. Early users are already split on how the models perform, but the licensing choice itself seems to echo the larger debate over openness, accountability and where AI development is headed.
And, perhaps most strikingly, OpenAI put out gpt-oss-120B and gpt-oss-20B as open-weight MoE reasoning models with that Apache-style licence. Whether you love the quality or side with the vocal critics, it’s the first time since GPT-2 that such large weights have been shared openly.
Finally -- and maybe most symbolically -- OpenAI released gpt-oss-120B and gpt-oss-20B, open-weight MoE reasoning models under an Apache 2.0-style license. Whatever you think of their quality (and early open-source users have been loud about their complaints), this is the first time since GPT-2 that OpenAI has put serious weights into the public commons. China's open-source wave goes mainstream If 2023-24 was about Llama and Mistral, 2025 belongs to China's open-weight ecosystem.
A study from MIT and Hugging Face found that China now slightly leads the U.S. in global open-model downloads, largely thanks to DeepSeek and Alibaba's Qwen family. Highlights: DeepSeek-R1 dropped in January as an open-source reasoning model rivaling OpenAI's o1, with MIT-licensed weights and a family of distilled smaller models.
VentureBeat has followed the story from its release to its cybersecurity impact to performance-tuned R1 variants. Kimi K2 Thinking from Moonshot, a "thinking" open-source model that reasons step-by-step with tools, very much in the o1/R1 mold, and is positioned as the best open reasoning model so far in the world. Z.ai shipped GLM-4.5 and GLM-4.5-Air as "agentic" models, open-sourcing base and hybrid reasoning variants on GitHub.
Baidu's ERNIE 4.5 family arrived as a fully open-sourced, multimodal MoE suite under Apache 2.0, including a 0.3B dense model and visual "Thinking" variants focused on charts, STEM, and tool use. Alibaba's Qwen3 line -- including Qwen3-Coder, large reasoning models, and the Qwen3-VL series released over the summer and fall months of 2025 -- continues to set a high bar for open weights in coding, translation, and multimodal reasoning, leading me to declare this past summer as " VentureBeat has been tracking these shifts, including Chinese math and reasoning models like Light-R1-32B and Weibo's tiny VibeThinker-1.5B, which beat DeepSeek baselines on shoestring training budgets.
OpenAI just put gpt-oss-120B and gpt-oss-20B out there under an Apache-2.0-style licence, and that feels like a pretty bold statement. These are open-weight mixture-of-experts models built for reasoning, showing up at a time when the AI scene resembles a patchwork of open and closed projects, big and tiny, Western and Chinese, cloud-hosted and on-prem. Some early users have already complained about the quality, so the reaction is far from unanimous.
Still, it’s the first occasion since GPT-2 that a big lab has released a model of this size with a permissive licence. I’m not sure yet whether developers will start building on it or drift toward other options. The move could point to a wider tilt toward openness, but the real effect will hinge on how the models handle everyday tasks and how fast the surrounding tools can adopt them.
Right now, the headline is the licence choice rather than any clear performance edge.
More in Policy & Regulation
Google Translate adds real‑time speech translation to any headphones
Google’s latest update to its translation service expands the reach of spoken‑language assistance...
After ditching AI fitness apps and a Fitbit, I return to Peloton classes
The surge of AI‑powered workout assistants has turned many home gyms into data‑driven command...
Trump signs executive order threatening to punish states that pass AI laws
President Donald Trump signed an executive order this week that threatens to withhold federal funds...
Congress Raises Alarm as Section 702 Expands Beyond Original Terror Surveillance
Seventeen years after Section 702 was written into law, the very framework that was meant to target...
NY signs two bills requiring ads to disclose AI‑generated avatars
New York’s governor put his signature on two pieces of legislation Thursday, targeting a niche...
Common Questions Answered
What are the names and sizes of the open-weight models OpenAI released under an Apache-2.0-style license?
OpenAI released two open-weight models: gpt-oss-120B, a 120‑billion‑parameter mixture‑of‑experts model, and gpt-oss-20B, a smaller 20‑billion‑parameter version. Both are designed for reasoning tasks and are now publicly available under an Apache‑2.0‑style license.
Why is the release of gpt-oss-120B and gpt-oss-20B considered a symbolic move for the AI community?
The release marks the first time since GPT‑2 that OpenAI has placed substantial model weights into the public commons, signaling a shift toward greater openness amid regulatory pressure for transparency. It also highlights the growing influence of China’s open‑source ecosystem, positioning 2025 as a year for open‑weight models.
What type of architecture do the gpt-oss models use, and what is their primary intended use?
Both gpt-oss‑120B and gpt-oss‑20B employ a mixture‑of‑experts (MoE) architecture, which allows different expert subnetworks to specialize in various reasoning tasks. Their primary purpose is to provide high‑quality reasoning capabilities while remaining openly accessible.
How have early adopters responded to the quality of the newly released gpt-oss models?
Early adopters have voiced complaints about the quality of the gpt‑oss models, suggesting that performance may not yet match expectations set by proprietary counterparts. Despite these concerns, the open‑weight release is still seen as a significant step toward broader community scrutiny and improvement.