At the end of March, OpenAI announced plans to release its first open-source language model since GPT-2 later this year. Gradually, information is emerging about this upcoming model through recent discussions with the AI developer community.
The initiative is being spearheaded by Aidan Clark, OpenAI’s Vice President of Research. Sources indicate the project remains in its preliminary stages, but OpenAI intends to launch it by early summer. Designed as a “reasoning” model, akin to OpenAI’s previous “o-series” AI systems, the new product aims to achieve top performance benchmarks among available open-source reasoning models.
Unlike many open-source alternatives, which often incorporate stringent licensing restrictions, OpenAI is reportedly exploring a highly flexible license arrangement with minimal limitations on commercial usage or modification. Models from competitors such as Meta’s Llama and Google’s Gemma have previously attracted criticism from some developers due to their restrictive terms, issues that OpenAI seeks to avoid with its forthcoming release.
As competitive pressure mounts from rival labs adopting more open approaches, notably China’s DeepSeek, OpenAI has faced growing demands from the community to make its technology more accessible. DeepSeek, for instance, has rapidly built a substantial global user base while attracting attention from notable regional investors. Similarly, Meta’s open-source Llama models surpassed one billion downloads, underscoring the success achievable through more accessible model strategies.
The forthcoming OpenAI model is described as a straightforward “text-in, text-out” system, intended to run efficiently on high-end consumer-grade hardware. Interestingly, OpenAI is said to be considering a configuration permitting developers to toggle its reasoning capabilities on or off. This design choice echoes recent innovations from companies like Anthropic. While reasoning modes enhance accuracy, the downside remains higher latency. Sources suggest that if community feedback proves positive, OpenAI could release additional models, potentially including smaller, more specialized variants.
OpenAI CEO Sam Altman previously acknowledged the company’s mixed history regarding open-source initiatives, stating publicly earlier this year that OpenAI needs a re-evaluated strategy toward open source. During a Reddit Q&A in January, Altman remarked that while his views are not universally shared across the organization, he personally believes OpenAI should revisit how openly it distributes its technologies, acknowledging it as a strategic necessity moving forward.
Altman also emphasized that this upcoming model would undergo rigorous safety testing and “red-team” exercises, given its potential wide-scale use and modification post-release. According to internal plans, OpenAI will provide a detailed model card along with the model, explicitly outlining the findings from both internal and external evaluations to document its safety profiles and performance metrics transparently.
Historically, OpenAI has faced criticism from some experts for insufficient safety testing procedures and incomplete disclosure regarding newer models’ risks. In addition, Altman himself came under scrutiny and faced leadership challenges at OpenAI in late 2023, partly due to disagreements and concerns raised surrounding the thoroughness of model safety assessments. OpenAI, however, seems intent this time on addressing those past controversies with greater transparency and caution.
OpenAI has not yet responded publicly regarding additional details of the new model or its impending launch timeline.
