OpenAI’s first open model in five years is coming
OpenAI is set to release an “open” AI model for free download in early summer, marking its first such release in about five years. The model is expected to outperform open models from Meta and DeepSeek.
According to TechCrunch, OpenAI leaders are considering a key feature that could make the open “reasoning” model highly competitive: the ability to connect to OpenAI’s cloud-hosted models to better answer complex queries. During a recent meeting with open-source AI community developers, CEO Sam Altman described this capability as a “handoff.”
If implemented, the feature will allow the open model to make calls to the OpenAI API, accessing larger models for significant computational power. The open model’s ability to access tools like web search and image generation is still unclear.
The idea for the handoff feature originated from a developer’s suggestion during one of OpenAI’s recent developer forums. OpenAI has been hosting community feedback events to shape its upcoming open model release.
A local model that taps into more powerful cloud systems is reminiscent of Apple Intelligence, which combines on-device models with models running in “private” data centers. This feature could generate incremental revenue and draw more open-source community members into OpenAI’s premium ecosystem.
Details on the handoff feature, including pricing and rate limits, remain unclear. The model is still in the early stages, and aspects of it may change or not be finalized.
OpenAI is training a new model from scratch for its open release, rather than repurposing an existing one. The open model is expected to underperform OpenAI’s o3 but outperform DeepSeek’s R1 reasoning model on certain benchmarks.