New Open models by OpenAI -- gpt‑oss‑120b and gpt‑oss‑20b, for developers Introduction OpenAI has officially launched two long-awaited ope...
New Open models by OpenAI -- gpt‑oss‑120b and gpt‑oss‑20b, for developers
Introduction
OpenAI has officially launched two long-awaited open-weight language models: gpt-oss-120b and gpt-oss-20b. These models are designed with developers in mind—providing accessible, high-performing LLMs that can be run and customized on local hardware, under a permissive Apache 2.0 license. This marks the most significant open-weight release from OpenAI since GPT-2 in 2019 and addresses developer demand for transparent, modifiable AI stacks.
Two Models for Flexible Deployment:
gpt-oss-120b: 117B parameters, production-grade, competitive with proprietary o3/o4-mini models, fits on a single H100 GPU. Ideal for server/enterprise applications requiring maximum reasoning and tool use.
gpt-oss-20b: 21B parameters, optimized for low hardware requirements, can run on consumer laptops/desktops with 16GB RAM, perfect for on-device and edge use
Key Features and Capabilities
- Open Weights: Both models make their underlying weights public, enabling transparency, deeper inspection, and easy fine-tuning for specific domains or applications.
- Agentic and Tool Use: Capable of function calling, web-browsing, Python code execution, and structured outputs. They can be connected to APIs or used as part of a broader agentic automation framework.
- Configurable Reasoning: Developers can adjust reasoning depth (“low”, “medium”, “high”) via system prompts, trading off speed for thoroughness as needed.
- Harmony Response Format: Models are trained to work best with OpenAI’s harmony format, enhancing instruction following and usability.
- Performance: On benchmarks like Codeforces and multi-subject exams, gpt-oss-120b nearly matches the performance of OpenAI’s latest commercial models, outperforming other open competitors like DeepSeek R1 and Qwen in most reasoning tasks.
- Fine-tuning Support: Both models are designed for full parameter fine-tuning; gpt-oss-120b supports this on GPU infrastructure, while gpt-oss-20b is suitable for experimentation on even modest consumer hardware.
Developer Impact
- Free and Flexible: Download models via Hugging Face or deploy through cloud partners (AWS, Azure) with minimal restrictions. License permits commercial use, redistribution, and integration into other projects.
- Privacy and Security Control: Run models locally for sensitive or private data applications, with no requirement for sending data to the cloud, and enforce your own security/safety policies as needed.
- Extensive Customization: Full support for fine-tuning and extension. Modify for domain-specific reasoning, compliance use cases, or language localization.
- Agentic Workflows: Integration-ready for custom AI agents, tools, and workflows—including direct interaction with browsers, databases, and developer tools.
Safety and Community Notes
- Risk Profile: Open weights increase the risk of misuse; OpenAI urges developers to add additional safeguards for applications with sensitive or critical outputs, as open models allow deeper modification and removal of default protections.
- Community-Driven: Models shaped with significant feedback from the open-source and research community, reinforcing OpenAI’s renewed commitment to openness and developer empowerment.
How to Get Started
Download:
Hugging Face: huggingface-cli download openai/gpt-oss-120b or openai/gpt-oss-20b
Pip install: pip install gpt-oss
- https://gpt-oss.com/ -- Click the link to test the models in a web as like a chatgpt.
- Run and Integrate: Use the CLI or Python API to run locally, connect to your apps, or fine-tune for custom workflows.
- Documentation: See the official Open AI site and Harmony Format guides for detailed usage, customization, and best practices
