Settings show up first in inference software
Settings for "Granite 5," an unannounced IBM language model, appear in a change proposal to "SGLang," public inference software for running AI models, @aitrackerbot, an X account that tracks the appearance of models, reported on Oct. 26 (the same day in Japan time). The proposal had been public since Sept. 25, according to the account.
According to @aitrackerbot, Granite 5 comes in two versions, with 20 billion and 120 billion parameters. Both use a mixture-of-experts design that switches between and uses only some of many expert components, with 56 and 128 experts, respectively. They are meant to run on Nvidia's H100 in a format with weights compressed to 4 bits (MXFP4), according to the account. ORIENS has not been able to confirm the contents of the proposal, and IBM has not made any announcement.
Open models for enterprises
Granite is IBM's series of open-weight models built for enterprises. In the previous generation, "Granite 4.1," a small 8 billion parameter model delivered performance on par with the 32 billion parameter model that had previously been the top of the line.
If the 120 billion parameter version is released, it could become the largest model in the series. The timing of the weights' release and the license are not known.
