AI service OpenAI unveils Gpt-oss model, offers reward for exposed weaknesses
OpenAI, the leading artificial intelligence research organization, has made a significant stride by releasing its first open-weight AI language model (LLM) named gpt-oss. The models, gpt-oss-120b and gpt-oss-20b, are now available for local download and offline use, marking a milestone in the organization's mission to democratize AI research.
Key Features of gpt-oss
The gpt-oss models, built on a Transformer-based architecture with a mixture-of-experts (MoE) system, boast several impressive features. They support long context windows of up to 128,000 tokens, making them ideal for extended conversations or document processing. Moreover, the models are equipped with multi-level chain-of-thought (CoT) reasoning—Low, Medium, and High— enabling them to tackle complex problems by breaking down the reasoning process.
Trained primarily on an English, text-only dataset focusing on STEM fields, coding, and general knowledge, the models have been deliberately filtered to exclude chemical, biological, radiological, and nuclear threat data for safety. OpenAI has also applied deliberative alignment and an instruction hierarchy post-training to improve the models' refusal of unsafe prompts and reduce prompt injection risk.
Performance and Capabilities
On benchmarks like HealthBench and multilingual evaluations (MMMLU in 14 languages), gpt-oss-120b performs on par with or better than some of OpenAI's previous models, demonstrating competitive reasoning and multilingual capabilities. The smaller gpt-oss-20b, despite being five times smaller, performs comparably or better than some of OpenAI's earlier models.
The models support tool use and agentic behavior, such as web browsing, coding assistance, and complex system interaction, but they are not multimodal—only supporting text input/output.
The gpt-oss Bounty Challenge
To ensure the robustness and safety of the gpt-oss models, OpenAI has launched a bounty challenge, encouraging the community to identify security issues, biases, or potential exploitation avenues. This initiative offers a total of $500,000 for the discovery of novel flaws and vulnerabilities.
Details of the bounty include open-sourcing the code and weights to foster transparency and broad testing. Researchers are incentivized to safely disclose vulnerabilities found during use, contributing to improving model alignment and safety.
A New Era for OpenAI
The release of gpt-oss marks OpenAI's first "open" release of an AI large language model, signifying a new era for the organization. While the bounty challenge does not apply to the larger gpt-oss-120b model, it underscores OpenAI's commitment to ensuring that openly available large language models are safe and reliable for developers and researchers.
OpenAI has previously released open models, such as Whisper, and has partnered with an unspecified entity for the launch of gpt-oss. The open source project huggingface's Transformers can be used for installing and running local LLMs, but there's no real tie-in to that platform.
The deadline for the bounty challenge submissions is August 26th, 2025. The gpt-oss-20b model can run within 16 GB of memory, making it accessible for a wide range of users. The larger and more capable gpt-oss-120b model requires 80 GB.
[1] OpenAI. (2025). gpt-oss: An open-weight AI language model family. Retrieved from https://openai.com/blog/gpt-oss [2] OpenAI. (2025). gpt-oss: A new era for open-weight AI language models. Retrieved from https://openai.com/blog/gpt-oss-features [3] OpenAI. (2025). gpt-oss: Benchmarks and evaluations. Retrieved from https://openai.com/blog/gpt-oss-benchmarks [4] OpenAI. (2025). gpt-oss: The bounty challenge. Retrieved from https://openai.com/blog/gpt-oss-bounty
- The gpt-oss models, made available by OpenAI, are not only built on artificial-intelligence technology but also incorporate open-source science in their Transformer-based architecture with a mixture-of-experts (MoE) system.
- To ensure the safety and reliability of the gpt-oss models, OpenAI has launched a bounty challenge, encouraging the science and technology community to identify potential issues, biases, or exploitation avenues, demonstrating their commitment to artificial-intelligence that is both open source and safe for developers and researchers.
