OpenAI has introduced two new open-weight language models designed for advanced reasoning, optimised to run on local hardware. The models are intended to provide performance levels similar to some of the company’s smaller proprietary reasoning models. This marks the first time OpenAI has released open models since GPT-2 in 2019.
Read: Samsung Galaxy Tab S10 FE Review: Mid-range makes sense
Open-weight models are distinct from open-source models. While open-source models provide full access to code and data, open-weight models make the trained parameters publicly accessible. This allows developers to analyse and fine-tune the models for specific tasks without needing the original training data. According to OpenAI co-founder Greg Brockman, this approach allows users to “run them locally” and “behind their own firewall.”
In a separate announcement, Amazon confirmed that OpenAI’s new open-weight models are now available on its Bedrock generative AI marketplace within Amazon Web Services (AWS). This is the first time an OpenAI model has been offered on Bedrock, a move that Bedrock Director of Product Atul Deo called a “great open-source option… for customers.” This new offering comes as AWS faces slowing growth compared to its rivals.
The open-weight and open-source AI landscape has become increasingly competitive. While Meta’s Llama models were once considered the leaders, Chinese company DeepSeek released a powerful, cost-effective reasoning model earlier this year, while Meta has struggled to deliver Llama 4.
See also
The two new OpenAI models are gpt-oss-120b and gpt-oss-20b. The larger model, gpt-oss-120b, can run on a single GPU, while gpt-oss-20b is small enough to run directly on a personal computer. OpenAI states that these models have similar performance to its proprietary o3-mini and o4-mini models, particularly excelling in coding, competitive math, and health-related questions.
The models were trained on a text-only dataset that focused on science, math, and coding. OpenAI did not release benchmarks comparing them to rival models like DeepSeek-R1.
