Towards the end of March, Openai announced its intention to release its first “open” language model since GPT-2 within this year. Now, details about that model are beginning to flow from company sessions with the AI developer community.
Sources tell TechCrunch that Aidan Clark, Vice President of Research at Openai, is a major development of an open model in its very early stages. Openai is targeting early summer releases and aims to make the model, which is an inference model that follows the line of the open O-Series model, the top benchmark among other open inference models.
Openai investigates the highly tolerant licenses for each TechCrunch source, a model with few usage or commercial restrictions. Open models like Llama and Google’s Gemma have been criticized by some in the community for imposing troubling requirements.
Openai faces increased pressure from rivals such as the Chinese AI Lab Deepseek, which took an open approach to launching models. In contrast to Openai’s strategy, these “open” competitors make the model available to the AI community for experimentation and, in some cases, commercialization.
It has proven to be a hugely successful strategy for some outfits. Meta, which invested heavily in the Lama family of open AI models, said in early March that Lama had won over 1 billion downloads. Meanwhile, Deepseek quickly garnered a large global user base and attracted the attention of domestic investors.
Sources tell TechCrunch that Openai intended to be an open model. This is “text-in, text-out” and runs on high-end consumer hardware, allowing developers to switch between “inference” models similar to the recently released inference models. (Inference can improve accuracy, but at the expense of increased latency.) If the launch is well received, OpenAI may follow up with additional models.
In a previous public comment, Openai CEO Sam Altman said he believes Openai is on the wrong side of history when it comes to opening up technology sourcing.
“I’m going to get a sense of another open source strategy (I personally think I need it),” Altman said in a January Reddit Q&A. “Not everyone at Openai shares this view, nor is it a top priority at the moment. We’ll create a better model (to the future), but we’ll continue to hold our lead from the previous year.”
Altman also states that Openai’s upcoming open models will be thoroughly red teamed and evaluated for safety. Sources have told TechCrunch that the company plans to release model cards for the model. This is a thorough technical report showing the results of OpenAI internal and external benchmarks and safety tests.
“(b) Before release, we evaluate this model according to the preparation framework, like other models,” Altman said in a post last month in X. “(a) we’ll do some additional work considering we know this model will change after release.”
Openai has been in a hurry to test safety on the recent models and has reportedly not released model cards for others, causing the rage of some AI ethicists. Altman has also been accused of misleading Openai executives of the model’s safety review before his brief expulsion in November 2023.
We contacted Openai for comments and will update this article if we respond.