OpenAI to Launch Its First Open-Source Language Model Since GPT-2

OpenAI plans to roll out an "open" version of its language model this year, enabling developers to run it on their own hardware.  

In a March 31 update on X, CEO Sam Altman announced that the company would release a "new open-weight language model with reasoning" in the coming months. However, before launch, OpenAI aims to gather feedback on how to make it as useful as possible.  

“We’ve considered this for a long time, but other priorities took precedence. Now, it feels like the right time,” Altman said, noting that this would be the first "open-weight" model since GPT-2 in 2019.  

He also mentioned that some key decisions are still pending, prompting OpenAI to host developer events to collect input and experiment with early prototypes.

Source: Sam Altman

An open-weight language model allows public access for usage, downloading, modification, and deployment, though it is not entirely open-source. This represents a departure from OpenAI’s previous models, such as GPT-3 and GPT-4, which remained fully closed.  

According to Altman, OpenAI will kick off developer events for GPT-2 in the coming weeks in San Francisco, where the company is headquartered. These events will later extend to Europe and the Asia-Pacific region.

Source: Steven Heidel

Altman expressed enthusiasm about the potential applications of the new model, stating, "We're excited to see what developers create and how large companies and governments leverage it when they prefer to run a model independently."

He also acknowledged the need for additional safeguards, given that the model is expected to be modified after release.

This marks OpenAI’s first "open-weight" language model since GPT-2, which was initially released in a limited capacity in February 2019 before being fully made available in November of the same year.

On Feb. 12, Altman revealed that OpenAI is gearing up to launch GPT-4.5 and GPT-5 in the near future.  

The AI industry is becoming increasingly competitive, with the introduction of DeepSeek—a ChatGPT-like model reportedly developed at significantly lower costs and in a shorter time frame.  

Alibaba Group debuted its latest open-source AI model on March 26, aimed at delivering cost-effective AI solutions, while Google rolled out Gemini 2.5, its newest experimental model, on March 25.  

Meanwhile, Meta CEO Mark Zuckerberg shared in a March 19 Threads update that the company’s AI model family, Llama, which was first introduced in February 2023, has now been downloaded over 1 billion times.