OpenAI releases first open models since GPT-2

OpenAI has released a model with open weights for the first time since the launch of GPT-2.

The new release, called GPT-OSS, is available under the Apache 2.0 license and can be downloaded by anyone via Ollama.

GPT-OSS comes in two versions:

• one with 20 billion parameters
• one with 120 billion parameters

Both are built using a Sparse MoE (Mixture of Experts) architecture — meaning only a portion of the model is activated during each query. This approach reduces hardware requirements and boosts efficiency.

The smaller version can run on consumer GPUs with 16 GB of VRAM, while the larger one needs a professional-grade accelerator with 80 GB of memory.

The model supports chain-of-thought reasoning that’s both transparent and controllable — users can adjust the depth of analysis and generation speed. It also supports tool use (Python, web search, etc.), similar to GPT-4o. However, multimodality (image and audio processing) is not included.

In benchmark tests, GPT-OSS-120B performs at a level close to GPT-4o-mini, and in some tasks, even rivals GPT-4o itself. On AIME 2024, it outperforms all others. Meanwhile, the smaller model delivers near-comparable results — with far more modest hardware needs.