OpenAI has made a significant move in the AI space with the release of two new open-weight models, GPT-OSS-120B and GPT-OSS-20B. These models, available for free download via Hugging Face under the permissive Apache 2.0 license, mark OpenAI’s first open model release since GPT-2 five years ago. This launch comes just before the anticipated release of GPT-5, signaling OpenAI’s return to more open practices in AI development.
GPT-OSS-120B vs 20B: Tailored for Different Users
The two newly released models serve different purposes:
- GPT-OSS-120B: This model is designed for deployment on a single Nvidia GPU, making it suitable for more robust computing environments.
- GPT-OSS-20B: This model is lightweight and can run on consumer-grade laptops with just 16GB of RAM, providing more accessibility for everyday users.
Both models are text-only and do not support multimodal tasks like image or audio generation. However, they can be used as intelligent agents, capable of interacting with more advanced OpenAI cloud models via APIs, creating hybrid workflows for users.
Powered by Mixture-of-Experts for Efficient Performance
OpenAI developed these models using a Mixture-of-Experts (MoE) architecture, activating only around 5.1 billion parameters per token in the 120B model for improved efficiency and speed. This approach ensures that the models perform well without unnecessary resource consumption. To further enhance their capabilities, high-compute reinforcement learning was applied during post-training, optimizing their reasoning abilities to match those of OpenAI’s proprietary models like the o-series.
Performance Benchmarks: Strong Results, But Room for Improvement
According to OpenAI, these open models perform well on several key benchmarks:
- Codeforces scores:
- GPT-OSS-120B: 2622
- GPT-OSS-20B: 2516
These scores outperform DeepSeek’s R1, though they still lag behind o3 and o4-mini.
However, a notable issue remains: hallucinations. On the PersonQA benchmark:
- GPT-OSS-120B hallucinated 49% of the time
- GPT-OSS-20B hallucinated 53%
This is significantly higher than the hallucination rates seen in OpenAI’s proprietary models like o1 (16%) and o4-mini (36%). OpenAI attributes these discrepancies to the smaller parameter activation in these open models and limited world knowledge, a necessary trade-off for greater efficiency.
Security and Licensing: Open but Careful
To address potential misuse, OpenAI has published a white paper alongside the launch, detailing their internal and external testing processes. While the models may slightly increase the capabilities of bad actors, OpenAI asserts that they do not pose a high risk for cyber or biochemical threats, even with fine-tuning.
One key aspect is that OpenAI has not released the training data for these models, likely due to ongoing legal issues regarding copyrighted content. Nevertheless, the Apache 2.0 license provides developers with full commercial rights, encouraging startups and enterprise users to adopt the models.
A Strategic Shift Back to Open Models
After years of keeping its models closed and proprietary, OpenAI is shifting back towards an open-source approach. This move is likely in response to the rise of competitors like China’s DeepSeek, Moonshot AI, and Alibaba’s Qwen, as well as Meta’s LLaMA, which previously led the open model space but has since lost momentum. OpenAI’s decision also comes amid political pressure in the U.S. to align AI innovation with democratic values. CEO Sam Altman recently admitted that OpenAI may have been “on the wrong side of history” regarding openness and transparency.
Final Thoughts
The release of GPT-OSS-120B and 20B is more than just a product launch—it’s a strategic pivot for OpenAI. With free commercial access, strong performance benchmarks, and a focus on agent-style tasks, these models are setting the stage for the future of open AI development. As GPT-5 approaches, OpenAI’s commitment to openness may well reshape the landscape of AI in the years to come.
