This video provides an initial review and comparative analysis of OpenAI's recently released open-source, Apache 2.0 licensed large language models: GPT-OSs-120B and GPT-OSs-20B. It highlights their design for agentic workflows with strong instruction following and tool use capabilities, noting their Mixture-of-Experts architecture and a 128K context length. Practical tests are conducted, including generating a browser-based operating system, a Python game, and a website, alongside a roleplay scenario. Surprisingly, the smaller 20B model often outperforms the larger 120B in terms of functionality and speed, and demonstrates a more permissive interpretation of content policies. The models are lauded for their clear, organized instructions and outputs, making them accessible for local deployment.
Exploring OpenAI's New Open-Source LLMs: GPT-OSs-120B and GPT-OSs-20B
0:00 OpenAI's First Open-Source Release Since GPT-2: OpenAI has released two new open-source, open-weight, Apache 2.0 licensed language models: GPT-OSs-120B and GPT-OSs-20B.
0:48 Model Capabilities and Hardware Requirements: The 120B model achieves near GPT-4 Mini performance on core reasoning and runs efficiently on a single 80GB GPU. The 20B variant delivers GPT-3 Mini-like results and can run on edge devices with just 16GB of memory.
1:34 Design for Agentic Workflows: Both models are designed for agentic workflows, featuring exceptional instruction following capabilities, including tool use (e.g., web search, Python code execution).
2:27 Architecture and Context Length: Both models offer a 128K context length and utilize a Mixture-of-Experts (MoE) architecture, which allows them to run on more constrained hardware despite their parameter count.
3:50 Accessibility and Deployment Options: The models are tested using Olama and Open Web UI. OpenAI provides comprehensive instructions for running these models with various methodologies like Transformers, VLM, PyTorch, Olama, and LM Studio, promoting accessibility.
5:08 Browser OS Generation (120B): The 120B model successfully generates a browser-based OS (HTML/CSS/JS) with acceptable speed. While initial app openings were behind the background, the model successfully fixed this issue on the second attempt, demonstrating functional, draggable, and resizeable windows.
10:48 Browser OS Generation (20B) - Unexpected Performance: The 20B model generates the OS significantly faster and, surprisingly, includes a functional terminal application that supports commands like 'echo', 'clear', and 'help', working correctly on the first try and outperforming the larger 120B model in this task.
17:50 Python Game Generation (20B): The 20B model successfully generates a functional retro-style Python Snake game with score tracking and basic movement, further showcasing its coding competency.
19:44 Python Game Generation (120B) - Functional Failure: In contrast, the 120B model's attempt to generate the Python game resulted in a non-functional script, highlighting a performance disparity where the smaller model excelled.
23:28 Website Generation (20B vs. 120B): Both models generate functional, mobile-responsive websites. The 20B creates a structured site with fake testimonials, while the 120B provides detailed customization instructions, strong hover effects, and a more intricate contact form with a dropdown.
28:50 Roleplay Test & Policy Interpretation: In a roleplay scenario, the 20B model engages enthusiastically, interpreting content policy more permissively for non-explicit content. The 120B model, however, strictly adheres to policy, refusing the request.
32:27 Overall Impressions and Future Potential: Both models are praised for their clear, organized, and instructive outputs, making them valuable educational tools. The 20B model stands out for its surprising performance, speed, and willingness to engage in tasks, suggesting significant future potential when integrated with tool use for agentic workflows.