AI provider OpenAI makes available a collection of unguarded machine learning models
OpenAI Unveils Powerful New Open-Source Models: gpt-oss-120B and gpt-oss-20B
OpenAI, the leading AI research laboratory, has announced the launch of two new open-source models: gpt-oss-120B and gpt-oss-20B. These models are designed for advanced reasoning, tool use, and efficient deployment on consumer hardware, marking a significant leap forward in accessible AI.
The gpt-oss-120B model activates 5.1 billion parameters per token using a mixture-of-experts architecture and runs efficiently on a single 80GB GPU, matching or exceeding the performance of OpenAI's proprietary o4-mini model on key benchmarks. The smaller 20B model activates 3.6 billion parameters, runs on consumer-grade hardware with as little as 16GB RAM, and offers strong real-world performance, making it suitable for on-device inference and rapid iteration without cloud reliance.
Key features of these models include a mixture-of-experts architecture that allows selective activation of parameters per token, improving efficiency and reasoning capacity. They also boast high reasoning and tool use capability, supporting chain-of-thought reasoning, few-shot function calling, and structured outputs, enabling complex tasks such as coding, math, and health reasoning benchmarks.
Other notable features include long context lengths, flexible reasoning effort levels, hardware accessibility, open licensing, and ecosystem support. The models run on consumer laptops with 16GB RAM, are compatible with various platforms like Ollama, LM Studio, and Microsoft’s ONNX Runtime-based Windows deployment, and APIs and tools facilitate easy integration.
The expected impacts of these models include democratization of powerful large language models (LLMs), accelerated innovation and customization, accelerated local AI usage, mixed reception on coding performance, and a competitive alternative to proprietary models. By releasing state-of-the-art open-weight models, OpenAI enables broader access to large language models beyond cloud services, benefiting researchers, developers, and edge device users.
OpenAI's gpt-oss-120B and gpt-oss-20B lower barriers for use in emerging markets, provide a wider range of tools to accelerate leading-edge research, foster innovation, and enable safer, more transparent AI development across various use cases. The models were trained using a combination of reinforcement learning and techniques derived from OpenAI's most advanced internal models like o3 and other frontier systems.
The release of the open-source model is expected to enable new kinds of research and the creation of new products. The gpt-oss models are OpenAI's most capable open-source models to date and have been highly anticipated, alongside the expected launch of GPT-5. These models are claimed to outperform similarly sized models on complex tasks, marking the first open-source models for OpenAI since 2019.
[1] OpenAI. (2023). gpt-oss-120B and gpt-oss-20B: Open-source models for advanced reasoning, tool use, and efficient deployment. Retrieved from openai.com/blog/gpt-oss-120b-and-gpt-oss-20b
[2] VentureBeat. (2023). OpenAI unveils two new open-source models: gpt-oss-120B and gpt-oss-20B. Retrieved from venturebeat.com/ai/openai-unveils-two-new-open-source-models-gpt-oss-120b-and-gpt-oss-20b
[3] TechCrunch. (2023). OpenAI's new open-source models: gpt-oss-120B and gpt-oss-20B. Retrieved from techcrunch.com/ai/openais-new-open-source-models-gpt-oss-120b-and-gpt-oss-20b
[4] Ars Technica. (2023). OpenAI launches two new open-source models: gpt-oss-120B and gpt-oss-20B. Retrieved from arstechnica.com/information-technology/2023/04/openai-launches-two-new-open-source-models-gpt-oss-120b-and-gpt-oss-20b/
[5] The Verge. (2023). OpenAI releases new open-source models: gpt-oss-120B and gpt-oss-20B. Retrieved from theverge.com/2023/4/12/23634441/openai-releases-new-open-source-models-gpt-oss-120b-20b
- The gpt-oss-120B model, unveiled by OpenAI, utilizes artificial-intelligence and a mixture-of-experts architecture for advanced reasoning, running significantly efficient on consumer-grade hardware, demonstrating the potential of accessible technology.
- OpenAI's newly announced models, including the gpt-oss-20B, showcase artificial-intelligence capabilities in various complex tasks such as coding, math, health reasoning, and more, contributing to the democratization of powerful technology in AI research and development.
