OpenAI has just launched two open-weight AI models — GPT-OSS-120B and GPT-OSS-20B marking its biggest open-source move since GPT-2 over five years ago. These models, which are freely available on Hugging Face, are designed to offer developers state-of-the-art reasoning capabilities while remaining lightweight enough to run on standard hardware.
The GPT-OSS-120B model can run on a single Nvidia GPU, while the lighter GPT-OSS-20B model works on consumer laptops with at least 16GB RAM.
Why This Launch Matters
This is a major shift for OpenAI, which has long favored closed-source development to power its commercial offerings like GPT-4. But with pressure mounting from fast-growing Chinese AI labs like DeepSeek, Qwen (Alibaba), and Moonshot AI, OpenAI is opening up again.
According to CEO Sam Altman, the company is going back to its roots:
“We are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.”
This launch also aligns with calls from the U.S. government urging AI developers to open up more technology in order to promote ethical, global AI adoption.
Performance Benchmarks: How GPT-OSS Stacks Up
OpenAI claims GPT-OSS-120B and 20B are now among the top-performing open models, based on widely used AI benchmarks:
- Codeforces (coding tasks):
- GPT-OSS-120B scored 2622
- GPT-OSS-20B scored 2516
- Both beat DeepSeek R1 but still trail OpenAI’s proprietary o3 and o4-mini models.
- Humanity’s Last Exam (reasoning questions):
- GPT-OSS-120B: 19%
- GPT-OSS-20B: 17.3%
- These scores outperform DeepSeek and Qwen models but fall short of o3.
However, the models do hallucinate more than OpenAI’s newer proprietary models. On OpenAI’s internal PersonQA test, GPT-OSS models gave incorrect responses nearly 50% of the time, compared to 16% for o1 and 36% for o4-mini.
How These Models Work
Both GPT-OSS models use a mixture-of-experts (MoE) architecture, meaning only a subset of the model’s full parameters are active during any given task — boosting efficiency. For example:
- GPT-OSS-120B has 117 billion total parameters,
but only 5.1 billion parameters are active per token.
They were trained using reinforcement learning (RL) techniques, the same approach used in OpenAI’s premium o-series. This allows them to think step-by-step and use tools like Python or web search in reasoning chains. However, these are text-only models — they can’t generate or analyze images, audio, or video.
Open License, But Not Fully Open Source
OpenAI is releasing the models under the Apache 2.0 license, a very permissive license that allows:
- Free commercial use
- No royalty or permission required
- Full model weights available to download
However, OpenAI will not release the training data, citing ongoing lawsuits over copyright and safety concerns.
Safety and Risks
OpenAI delayed the release several times to ensure safety. Tests showed GPT-OSS could marginally increase biological risk capabilities when fine-tuned, but not enough to reach dangerous thresholds. According to the company, external evaluations showed no evidence of extreme misuse potential for now.
What’s Next?
While GPT-OSS raises the bar for open-weight AI, competitors aren’t staying quiet. Developers are now looking forward to:
- DeepSeek R2 (China’s upcoming model)
- Meta’s Superintelligence Lab open model
These will likely push the performance of open models even further in the coming months.
With GPT-OSS, OpenAI is stepping back into the open-source space — not just for developer goodwill, but also to keep up with global AI competition. While these models don’t beat the latest GPT models, they provide free, powerful tools for developers, startups, and researchers looking to build on open infrastructure.
