Bel
A rumored internal codename at OpenAI for a next-generation foundation model. It is said to exceed 10 trillion parameters and may adopt a Mixture-of-Experts (MoE) architecture. It is positioned as a foundation model succeeding GPT-6 and as a candidate for the threshold of artificial general intelligence (AGI).
Timeline (last 90 days)
A total of 10T parameters most likely corresponds to an MoE model with activated parameters in the hundreds of billions, rather than a dense model.
The leak claims that Bel's total parameter count is reportedly over 10 trillion (10T).
OpenAI internally views Bel as a foundation model toward the AGI threshold.
The original post mentioned that Bel is Doug's successor, where Doug most likely refers to some model project or codename inside OpenAI.
All current information about Bel comes from this single Reddit post, with no second independent source to corroborate it.
A leak post about Bel, a new internal model at OpenAI, appeared on Reddit.
Information about the 'Bel' model comes from community leak channels and has not yet been confirmed by OpenAI officially.
In the context of large models, '10T' may also refer to the total number of tokens in training data rather than the parameter count.
Whether the codename 'Bel' corresponds to the externally anticipated GPT-5 remains undetermined.
Reddit community leaker Leo claimed that OpenAI has just completed its next pre-trained model exceeding 10 trillion parameters in scale, codenamed 'Bel'.
All Facts (7)
OpenAI internally views Bel as a foundation model toward the AGI threshold.
50%UnverifiedAll current information about Bel comes from this single Reddit post, with no second independent source to corroborate it.
50%UnverifiedA total of 10T parameters most likely corresponds to an MoE model with activated parameters in the hundreds of billions, rather than a dense model.
50%UnverifiedThe leak claims that Bel's total parameter count is reportedly over 10 trillion (10T).
50%UnverifiedThe original post mentioned that Bel is Doug's successor, where Doug most likely refers to some model project or codename inside OpenAI.
50%UnverifiedIn the context of large models, '10T' may also refer to the total number of tokens in training data rather than the parameter count.
50%UnverifiedA leak post about Bel, a new internal model at OpenAI, appeared on Reddit.
50%