
After more than five years since the release of GPT-2, OpenAI has unveiled two new open-weight models GPT OSS-120B and GPT OSS-20B. This marks a significant shift for the company, which had long favored a closed-source approach. These models are available under the permissive Apache 2.0 license, making them accessible to developers and enterprises without restrictive licensing obligations.
Unlike their multimodal counterparts, these models are strictly text-based. The larger GPT OSS-120B is optimized for a single Nvidia GPU setup, while the smaller GPT OSS-20B can operate on consumer laptops with just 16GB RAM. Both models support agent-style tasks and complex reasoning workflows, although they lack capabilities for processing images or audio. However, they can route such queries to more advanced OpenAI models via cloud APIs, functioning as intelligent intermediaries.
The models utilize a Mixture-of-Experts (MoE) architecture, activating only around 5.1 billion parameters per token in the larger version. This structure enhances responsiveness while minimizing compute demands. Moreover, post-training with high-compute reinforcement learning boosts their reasoning abilities, aligning them closely with OpenAI’s o-series models.
Performance Metrics, Limitations, and Efficiency Trade-offs
Performance-wise, GPT OSS-120B scored 2622 on Codeforces, while the 20B version reached 2516. Both scores outperform those of competing models like DeepSeek’s R1. Nevertheless, these open models still lag behind OpenAI’s o3 and o4-mini when it comes to raw capabilities.
Despite their strengths, hallucination remains a concern. According to benchmark results, the 120B model hallucinated 49% of the time, and the 20B model did so 53% of the time. This contrasts sharply with the lower hallucination rates in OpenAI’s older proprietary models. The increased rate is attributed to limited world knowledge and fewer activated parameters expected drawbacks when prioritizing efficiency over depth.
Yet, these limitations may not outweigh the benefits for many developers. With free access under Apache 2.0, these models offer practical solutions for businesses aiming to deploy capable AI without incurring high licensing costs. This opens new doors for experimentation, especially for smaller players in the AI ecosystem.
A Strategic Pivot Amid Global AI Developments
OpenAI’s reentry into open-weight model development is not just a technical shift it’s also a strategic response to growing global competition. Recently, organizations in China such as DeepSeek, Moonshot AI, and Alibaba’s Qwen have rapidly advanced in the open-source AI domain. Meanwhile, earlier leaders like Meta’s Llama have seen their momentum slow.
Although OpenAI has not disclosed the training datasets for GPT OSS, the company has evaluated risks associated with potential misuse. Internal and third-party assessments concluded that these models, while powerful, do not cross critical risk thresholds.
The timing of this move aligns with increasing governmental calls particularly from the U.S. to democratize AI technology through open-source efforts. By offering powerful yet accessible models, OpenAI aims to re-establish its relevance among developers, researchers, and institutions seeking transparent and flexible AI tools.