IBM Granite Speech hits 12,600 RTFx

IBM released Granite 4.2 in 3B, 8B, and 30B sizes alongside a 470M-parameter Granite Speech 5.0 model that reached 12,600 RTFx on one H200—enough to transcribe three hours of audio in a second—in IBM’s launch report. Granite 4.2 uses an Apache 2.0 license and adds thinking and tool-calling modes, while the speech model drops the LLM backbone for efficient streaming transcription. Builders should test the two as separate deployment choices, extending the archive’s case for shrinking open models to fit real inference budgets.