OpenAI Releases Free AI Models That Run Locally - How Do They Compare to GPT-4?

OpenAI has announced two new language models, gpt-oss-120b and gpt-oss-20b that come very near to matching the performance of GPT-4 with one big difference – the new models are entirely open-weight models that are free to use and anyone can use, change or run on their own computer. These are free to use models that run under the Apache 2.0 license. They give developers and researchers real control over powerful AI tools. 

GPT-4 and its smaller version - o4-mini doesn’t let developers do so because it is closed and only has an API. OpenAI believes that GPT-OSS-120B performs better than GPT-o4-mini model on a lot of real-world tests, such as HealthBench for healthcare, AIME for arithmetic, and general reasoning tasks. 

Anyone can download, use, and tweak GPT-OSS-120B for free, unlike GPT-4 and GPT-4o, which are still behind an API paywall.  Even while the smaller GPT-OSS-20B is not as strong as GPT-4, it still exceeds many commercial models at some tasks. It can also run on devices with as low as 16 GB of RAM, which makes it very useful for lightweight apps and AI on devices.

How Are New Models Different from GPT-4

Access and customisation: GPT-4 (including o4-mini) is still closed.  There are no limits on how you can use or change GPT-OSS models. 

Hardware Requirements: GPT-4 works in the cloud. You only need one 80 GB GPU to run GPT-OSS-120B, and you only need 16 GB to operate GPT-OSS-20B.  This allows developers more freedom when deploying locally.

Openness: OpenAI has made the architecture specs for the OSS models public.  Most of the time, the public doesn't know much about how GPT-4 works.

Safety: Both model lines are tested for safety, but with GPT-OSS, OpenAI is encouraging outside people to audit and experiment with the models by distributing them as-is.

Reasoning Tools: GPT-OSS models can do things like chain-of-thought, tool use, and function calling, which were only available in GPT-4 class models before. 

Why It Matters

This is the first time OpenAI has made a model that is almost as powerful as GPT-4 and let the public use it freely.  gpt-oss-120b is a viable alternative to GPT-4 for people who are looking for openness, free experimentation, and the possibility to run AI models without having to use cloud infrastructure.  This might be a big change for developers, AI researchers, and corporations that have unique privacy demands.  It implies fewer restrictions, more options for customisation, and less vendor lock-in, all while still getting results that are almost as good as GPT-4. 

Read More: OpenAI Fights to Limit NYT’s Access to 120 Million ChatGPT Logs