# Building with Open Models

Source: https://www.youtube.com/watch?v=1HL2YHRj270
Recap page: https://rapidrecap.app/video/1HL2YHRj270
Generated: 2025-10-08T17:36:09.356+00:00

---
## 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.

![Screenshot at 00:06: Dominik Kundell on stage at OpenAI DevDay 2025, introducing the session titled 'Building with Open Models' to an assembled audience.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-00-06.png)

**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 'generate_visualization' 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.

### GPT-OSS Model Overview

- gpt-oss-20b requires >= 16 GB VRAM
- gpt-oss-120b requires 80 GB VRAM
- Both are Apache-2 licensed for commercial use and fine-tuning.

### Key Capabilities

- Models feature reasoning, tool calling (including web browsing/Python interpreter), and custom fine-tuning.

### Performance Benchmarks

- GPT-OSS-20b performs on par with GPT-4 (O3) and GPT-4-mini on 'Humanity's Last Exam' (53.5% accuracy).

### Local Execution Demo

- Live demonstration showed gpt-oss-120b running locally on a MacBook Pro using an MPC server proxy to read financial data and execute Python code via the interpreter tool.

### Agent Customization

- Agents can be configured with custom tools (e.g., 'generate_visualization') using the Agents SDK, allowing for hybrid deployment strategies.

### Reasons for Open Models

- To ensure data privacy, low latency, and maximum flexibility for developers who need to run models locally or fine-tune them for specific tasks.

![Screenshot at 00:01: OpenAI DevDay \[2025\] title slide, setting the context for the presentation.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-00-01.png)
![Screenshot at 00:08: Dominik Kundell introducing the session topic: 'Building with Open Models', with audience in attendance.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-00-08.png)
![Screenshot at 00:48: Slide defining GPT-OSS, showing the two model sizes \(20b and 120b\) and their VRAM requirements.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-00-48.png)
![Screenshot at 03:02: Slide titled 'Humanity's Last Exam' comparing GPT-OSS performance against O3, O4-mini, and other models using accuracy percentage.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-03-02.png)
![Screenshot at 02:10: Slide detailing key features of the open models: Reasoning models, Tool calling as part of CoT, and Apache-2 licensing.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-02-10.png)
![Screenshot at 04:27: Slide titled 'Intelligence vs. active parameters' showing a scatter plot comparison of various open models against intelligence score and active parameters.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-04-27.png)
![Screenshot at 05:30: Slide outlining five key use cases for open models: Controlling robots, Discussing sensitive topics, Fine-tuning for expertise, Offline coding assistants, and Coding for on-premise cyber security needs.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-05-30.png)
![Screenshot at 07:58: Demonstration of running the gpt-oss-120b model locally via the command line, showing the server starting up.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-07-58.png)
![Screenshot at 08:17: Code snippet showing the agent configuration using the local responses API proxy and configuring the 'code\_interpreter\_tool'.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-08-17.png)
![Screenshot at 11:51: Demonstration of the agent running locally, accessing and processing files from the local file system without internet connectivity.](https://ss.rapidrecap.app/screens/1HL2YHRj270/00-11-51.png)