Today, OpenAI announced the launch of gpt-oss-120B and gpt-oss-20B, the company’s first open-weight language models since GPT-2, which debuted back in 2019. Unlike the proprietary models that power ChatGPT and API services, these new releases can be downloaded and run locally, giving developers full access to the underlying architecture.
What Makes gpt-oss Different?
Both models — the 120 billion and 20 billion parameter variants — are built with transformer-based architecturesand include support for simulated reasoning, tool usage, and deep prompt-level customization. OpenAI has introduced a flexible “chain of thought” (CoT) configuration, which lets users select between low, medium, or high reasoning levels by setting a single line in the system prompt. Higher settings use more resources but provide significantly improved output quality.
Designed with accessibility in mind, the models are tailored for different hardware capabilities:
- gpt-oss-20B uses a mixture-of-experts (MoE) architecture to reduce per-token usage to 3.6 billion parameters, making it feasible to run on consumer hardware with 16GB+ memory.
- gpt-oss-120B, while much more powerful, requires around 80GB of memory, ideally on AI-optimized hardware like Nvidia’s H100 GPU. It uses MoE to bring per-token usage down to 5.1 billion parameters.
Both support a 128,000-token context window, offering high flexibility for complex tasks.
“These models aren’t meant to replace our proprietary GPT versions,” an OpenAI spokesperson said.
“They’re built for users who need greater control, local processing, and more flexibility — all while staying within the OpenAI ecosystem.”
Performance Benchmarks & Limitations
While not on par with the upcoming GPT-5 or even GPT-4, gpt-oss-120B performs competitively, benchmarking between OpenAI’s proprietary o3 and o4-mini models. The smaller model trails slightly but performs admirably in tasks like math and coding.
In the Humanity’s Last Exam benchmark — a widely cited test of model comprehension and performance — the larger gpt-oss-120B scored 19%, compared to 24.9% for o3 and 34.8% for Google’s Gemini Deep Think. Clearly, the top-tier results are still held by proprietary models, but OpenAI’s offering is closing the gap for open-weight alternatives.
That said, there are some limitations: these models are text-only and do not support multimodality out of the box. They also don’t match the scale or fine-tuning polish of GPT-4 or ChatGPT Enterprise. However, for many developers and enterprises, the tradeoff for local control, speed, and data privacy is well worth it.
Why OpenAI Released These Now
The decision to release open-weight models comes as organizations increasingly seek AI solutions they can run on their own infrastructure — either for latency advantages, compliance reasons, or sensitive data handling.
Until now, many of those needs were met by third-party open models like Meta’s LLaMA or Mistral. With gpt-oss, OpenAI aims to provide a first-party solution that fits into its existing ecosystem, enabling hybrid workflows where proprietary and open models work side-by-side.
Both models are governed by the Apache 2.0 license, allowing for extensive customization, research, and commercial use.
Safety Testing and Alignment
OpenAI also addressed the potential for misuse — a major concern with powerful open models. According to the company, the new models were subjected to “worst-case scenario” testing, where teams deliberately attempted to tune them for harmful behavior.
“Even when optimized for bad outputs, the models underperformed in malicious tasks,” OpenAI said in its internal safety report.
“This is a result of deliberate alignment layers and instruction hierarchies designed to minimize abuse risk.”
In other words, while the models are open, they’re still built with safety and responsible use in mind — something OpenAI says is critical for long-term trust in the AI space.
How to Get Started
Both gpt-oss-120B and gpt-oss-20B are now available for download:
- Hugging Face
- GitHub
In addition, OpenAI is hosting the stock versions on its own infrastructure, so developers can experiment without needing to provision local hardware immediately. Detailed model cards and a technical research blog are also available for those interested in performance details and use case examples.
Final Word
With the release of gpt-oss, OpenAI is signaling a broader commitment to transparency, developer empowerment, and hybrid AI infrastructure. While these models don’t dethrone GPT-4, they’re a valuable resource for teams looking to deploy AI locally — or simply understand it more deeply.
Whether you’re an enterprise building privacy-first solutions, or a researcher fine-tuning models for niche tasks, OpenAI’s latest release brings open-weight AI back into the spotlight.
