Building with Open Models

Quick Overview

OpenAI released the GPT-OSS model series (GPT-OSS-20b and GPT-OSS-120b) under the Apache-2 license to offer developers high-performing, locally runnable large language models that support tool use, Chain-of-Thought reasoning, and custom fine-tuning, aiming to provide flexibility for use cases requiring on-premise deployment or data privacy.

Key Points: OpenAI released the GPT-OSS model family, consisting of gpt-oss-20b (requiring at least 16 GB VRAM) and gpt-oss-120b (requiring 80 GB VRAM). Both models are licensed under Apache-2, allowing for commercial use and fine-tuning. These open models support Chain-of-Thought reasoning and tool calling, including web browsing and Python interpreter tools. GPT-OSS-20b performs comparably to OpenAI's O3 and O4-mini models on 'Humanity's Last Exam' benchmark. The models are designed for local or on-premise execution, supporting privacy and low-latency use cases. The presentation included a live demonstration of the 120b model running locally on a MacBook Pro to manage financial data and generate visualizations. Developers can find resources and code on the GitHub repository linked during the presentation.

Context: The video captures a segment from the OpenAI DevDay 2025 keynote, where Dominik Kundell from Developer Experience at OpenAI introduces the company's new family of open-source large language models, named GPT-OSS. The presentation focuses on the capabilities and deployment flexibility of these new models, contrasting them with proprietary offerings and demonstrating their ability to use tools and run locally.

Detailed Analysis

Dominik Kundell announced the release of the GPT-OSS model series, comprising gpt-oss-20b and gpt-oss-120b, both released under the permissive Apache-2 license. The 20b model requires at least 16 GB of VRAM to run on consumer hardware like high-end desktops or MacBooks, while the 120b model requires 80 GB of VRAM, suitable for data centers or high-end desktops. A key feature highlighted is that these models support Chain-of-Thought reasoning and tool calling, including the ability to use Python code interpreters and web browsing tools, which is noted as a capability that some proprietary models like GPT-5 still struggle with or require specific training for. Benchmarking against 'Humanity's Last Exam' showed that gpt-oss-20b performs comparably to models like GPT-4 (O3) and GPT-4-mini (O4-mini). Kundell demonstrated the local execution capabilities by running the 120b model on his MacBook Pro via an MPC server proxy, successfully using the file system tool to read financial documents and the code interpreter tool to calculate portfolio growth, proving that powerful models can run entirely locally for enhanced privacy and low latency. He also showed how to equip agents with custom tools, like a 'generatevisualization' tool, demonstrating hybrid flexibility where local models handle sensitive tasks while leveraging larger models like GPT-5 when necessary. The presentation concluded by directing attendees to the GitHub repository for further exploration.