Sam Altman announced that OpenAI is set to release an open-weight artificial intelligence model during the upcoming summer. He expressed excitement about this powerful new language model on X, indicating it will feature reasoning capabilities.
This strategic decision comes in response to the notable success of DeepSeek’s R1 model and the popularity of Meta’s Llama models. Altman previously mentioned that OpenAI was "on the wrong side of history" regarding open-source models, suggesting a shift in their direction.
The push for an open-weight model is significant, as it presents an opportunity for OpenAI to demonstrate cost-effective training methods, especially considering that DeepSeek reportedly trained its model at a lower cost than many other large AI systems. Clement Delangue, cofounder and CEO of HuggingFace, praised this development, highlighting the increasing recognition of open-weight models’ potential.
OpenAI currently provides its AI through a chatbot and cloud services, while competitors like DeepSeek and others allow users to download and modify their open-weight models. Open-weight models are advantageous for tailored applications, especially in handling confidential information due to their lower usage costs.
In additional updates, Steven Heidel, a technical staff member at OpenAI, confirmed that users would be able to run the upcoming model on their hardware. AI safety researcher Johannes Heidecke emphasized that OpenAI would conduct strict testing to mitigate risks associated with open-weight models, addressing concerns about possible misuse for malicious purposes. The company maintains a commitment not to release models that pose catastrophic risks.
Furthermore, OpenAI has initiated a call for developers to apply for early access to the forthcoming model and plans to hold events showcasing early prototypes in the weeks ahead.
As Meta pioneered the open-source approach with its Llama models, there is a growing trend within the industry towards open-weight AI, although some researchers critique the level of transparency provided by these models regarding their training data.
