Open AI is now finally Open AI.
OpenAI has just released two new advanced open-weight models, gpt-oss-120b and gpt-oss-20b.
These models are designed for high-end reasoning tasks and offer flexibility for customization and deployment.
Sam Altman is very bullish about this.
What is an Open-Weight Model?
An open-weight model is an AI, often an LLM, with publicly shared parameters or "weights" learned during training.
Sharing them allows users to easily run, explore, and customize the model on their own hardware, unlike closed models, which are only accessible through an API with proprietary weights.
Two Model Sizes
gpt-oss-120b: Powerful, large model with 120 billion parameters, specially optimized for smooth performance in data centers and on high-end desktops and laptops.
gpt-oss-20b: Versatile, medium-sized model with 20 billion parameters designed to be accessible to many, efficiently running on most modern desktops and laptops.
To try out these models, you can visit: https://gpt-oss.com/
Apache 2.0 License
Both models are provided under the Apache 2.0 license, so you can easily use, modify, and share them. This opens up exciting possibilities for us to create, experiment, personalize, and launch commercial applications without worrying about copyleft requirements or patent issues.
GitHub repository: https://github.com/openai/gpt-oss.
Designed for Agentic Tasks
The claim is that these models are really good at handling complex, multi-step tasks easily and that they shine in following instructions and make great use of their impressive tools, such as web searches and Python coding, to help solve problems more effectively.
Deep Customization & Fine-Tuning
Users have granular control over the models' performance. The reasoning effort can be adjusted to low, medium, or high, balancing performance with resource consumption.
The models support full-parameter fine-tuning, allowing developers to adapt them precisely to specific use cases and achieve optimal results.
This sounds great, but of course, it needs to be tested.
Model Performance
The models' performance seems promising. 📈
| Benchmark | gpt-oss-120b | gpt-oss-20b | OpenAI o3 | OpenAI o4-mini |
|---|---|---|---|---|
| Reasoning & Knowledge | ||||
| MMLU | 90.0 | 85.3 | 93.4 | 93.0 |
| GPQA Diamond | 80.1 | 71.5 | 83.3 | 81.4 |
| Humanity's Last Exam | 19.0 | 17.3 | 24.9 | 17.7 |
| Competition math | ||||
| AIME 2024 | 96.6 | 96.0 | 95.2 | 98.7 |
| AIME 2025 | 97.9 | 98.7 | 98.4 | 99.5 |
Research blog: https://openai.com/index/introducing-gpt-oss/
Chain-of-Thought (CoT) Access
OpenAI now provides complete access to the model's reasoning process.
Finally, some transparency. That's great for easier debugging, understanding how the model arrived at an output, and building higher trust in its results.
Official Blog Post: https://openai.com/index/introducing-gpt-oss/
Conclusion
I hope this overview of OpenAI's new open-weight reasoning models has been helpful. These models represent a significant step in making advanced AI more accessible and customizable for developers and researchers.
If you have any questions or need further information, comment below or contact me directly.
You can find me here: Francesco Ciulla
Top comments (3)
Thanks Francesco!
Thanks Eleftheria!
The release of GPT‑OSS‑20B and 120B under Apache 2.0 is a huge win for developers, researchers, and startups. Running powerful models locally, customizing them freely, and avoiding cloud lock-in? That’s the kind of openness we’ve been waiting for.
The benchmarks are solid, and the chain-of-thought access is a game changer for debugging and transparency.
I’ve already started experimenting with GPT‑OSS‑20B on my laptop smooth performance and surprisingly capable.
Excited to see how the community builds around this!