🚨 OpenAI Releases GPT‑OSS: Fully Open Models Up to 120B Parameters

Date: August 5, 2025

OpenAI has just shaken the AI world again — this time, by launching GPT‑OSS, a family of fully open-weight models with parameter counts up to 120B. This is OpenAI's first open-weight release since GPT‑2 in 2019, and it's already trending across developer communities and platforms.

What Is GPT‑OSS?

GPT‑OSS comes in two major variants:

gpt‑oss‑120B: A high-performance, mixture-of-experts model that can run efficiently on a single 80 GB GPU, matching OpenAI’s o4‑mini in benchmark quality.
gpt‑oss‑20B: A smaller, yet powerful version optimized for systems with just 16 GB memory, delivering performance similar to o3‑mini.

Both models are released under the Apache 2.0 license, meaning developers can freely download, fine-tune, and deploy them for commercial or personal use — a significant shift in OpenAI’s previously closed-weight policy.

Key Technical Highlights

Mixture of Experts (MoE): GPT‑OSS 120B uses 128 experts per layer, but only 4 are active per token, making inference both efficient and powerful.
Supports Chain-of-Thought, Tool Use, and Structured Output: These models handle reasoning tasks and complex tool interaction workflows exceptionally well.
High Safety Standards: OpenAI evaluated GPT‑OSS using its internal Preparedness Framework with external review, ensuring the model is safe for public release.

Partnerships & Deployments

OpenAI also partnered with platforms including:

Hugging Face
Amazon Bedrock & SageMaker
vLLM
llama.cpp, Ollama, and others

Amazon has already announced GPT‑OSS availability via Amazon Bedrock, claiming up to 3× better cost efficiency than Meta’s Gemini and 5× better than DeepSeek-R1.

Why This Matters

With GPT‑OSS, OpenAI is empowering developers who prioritize:

On-device inference
Cost control and privacy
Custom fine-tuning and deployment freedom

This move puts OpenAI in direct competition with other open-weight models like Meta’s LLaMA and Mistral, marking a major strategic pivot back toward openness.

Official Announcements & References

Conclusion

This release is not just about performance — it's about giving control back to developers. OpenAI’s GPT‑OSS could redefine the balance between open AI research and commercial productization. If you’ve been waiting for powerful, open, production-ready models to use in your apps, this is your moment.

Stay tuned on Tech Ahmed for benchmarks, tutorials, and deployment guides coming soon!