OpenAI Finally Opens Up
The newly-released GPT OSS models have finally put the "open" in OpenAI, but the company's flagship models are still under lock and key.
It is a great time to be working at a company like OpenAI. Recent advances in artificial intelligence (AI) are widely believed to be every bit as important as the rise of the internet was 30 years ago, and OpenAI is at the epicenter of many of these innovations. But in one regard, employees of OpenAI are deserving of our pity — by now they must be awfully tired of being razzed by literally everyone for working at a company with “open” in the name that is anything but open.
The tide may at last be beginning to turn, as OpenAI has just released a pair of open-source models called GPT OSS. Whether this release was spurred on by a desire to live up to the spirit of their name, or because competition is closing in on all sides, will have to be decided by the reader. But in any case, GPT OSS is not among the company’s flagship models like GPT-4.1 or o3. The released models are decidedly smaller, weighing in at roughly 20 and 120 billion parameters.
For the serious AI researcher, this slimmed-down release may be somewhat of a disappointment. But for a small academic lab or the weekend hobbyist, these smaller models are pretty exciting. The 120B model is on the heavier side, as it requires an NVIDIA H100 GPU with 80GB of memory for proper operation. These GPUs typically sell for tens of thousands of dollars — if you can find one for sale at all, that is. The 20B model, on the other hand, only requires 16GB of memory, so much more modest hardware is sufficient to get in on the action.
The new models are not going to approach the performance of OpenAI’s latest proprietary models, but benchmarks do show some respectable numbers. For the tinkerer or the small business that wants more privacy, GPT OSS could fit their needs quite well. The models can also be fine-tuned to meet the particular needs of an individual or organization.
The models are available now at Hugging Face, and tools like llama.cpp and Ollama have inference implementations ready to go. OpenAI has also stood up a playground where you can try the models out first without installing anything locally. This may be a small first step in the direction of open source AI models, but we can hope that the trend will continue with a major player like OpenAI dipping their toes in the water.
R&D, creativity, and building the next big thing you never knew you wanted are my specialties.