OpenAI's gpt-oss: Open Models for Custom AI Solutions

Dominik Kundel, from Developer Experience at OpenAI, presented gpt-oss, OpenAI's latest open model series, emphasizing its adaptability and seamless integration with GPT-5 for flexible and impactful AI development. "Open models are not always an option for you," Kundel stated, highlighting the necessity for customizable AI solutions.

Kundel spoke at OpenAI DevDay about gpt-oss, covering its core features, potential applications, and integration within the broader OpenAI ecosystem. This launch addresses the need for adaptable AI, especially in situations where data privacy, specialized hardware, or offline functionality are paramount.

The gpt-oss family comprises two models: gpt-oss-20b, suitable for high-end consumer hardware with at least 16 GB of VRAM, and gpt-oss-120b, designed for data centers and high-end desktops with 80 GB of VRAM. This range ensures accessibility across diverse hardware setups.

"We know that proprietary hosted models aren't always an option for you," Kundel explained. This recognition of the limitations of proprietary models underpins OpenAI's commitment to providing developers with choices that align with their unique constraints.

Kundel emphasized that gpt-oss models are reasoning models, permitting tool calling as part of the Chain-of-Thought (CoT) process and are licensed under Apache-2. This allows use in both commercial and research contexts, providing flexibility for developers.

"Both of the open models are reasoning models with variable levels of reasoning and raw chain of thought access," said Kundel. The Chain-of-Thought access is a key aspect, allowing users to understand the reasoning process behind the AI's conclusions.

According to Kundel, gpt-oss models are the only open models that can perform tool calling as part of the chain of thought, including web browsing and Python tool calling. This unique capability allows the model to combine tools and reason between them to achieve complex tasks.

He further noted, "These models should also instantly feel familiar to you if you've used our other reasoning models like OpenAI's o3, o4 mini or GPT-5." The intention is to provide a consistent developer experience across OpenAI's diverse range of models.

Kundel demonstrated gpt-oss capabilities live, including the ability to perform complex mathematical calculations, browse the web, and control robots. These capabilities underscore the model's versatility and potential applications.

Related Reading

• OpenAI's Frontier: Compute Scarcity, Proactive AI, and a Reshaped Human Experience
• OpenAI's Agentic Leap: Reshaping Software Development with Powerful APIs
• Sam Altman on Vertical Integration, Societal Co-evolution, and AI as the New Scientist

During the demonstration, it was revealed that even a complex task, like creating a custom visualization, could be delegated to GPT-5 by providing it as a tool for gpt-oss. This highlights the ability to use GPT-5 for tasks that gpt-oss may be less adept at, creating a hybrid system.

"That is real state-of-the-art level intelligence that you can run now locally on your computer," Kundel concluded, underscoring the impact of gpt-oss in bringing advanced AI capabilities to local environments.