IBM's Granite 4.2 agents trained in sandboxes
ibm released granite 4.2 models (3b, 8b, 30b) with 512k context and 'agentic rl' under apache 2.0. this raises the bar for open-source agents that can execute code and tools.
ibm launched the granite 4.2 family in 3b, 8b, and 30b sizes, featuring a 512k context window and a mode to toggle between 'thinking' and 'low-effort' inference. trained on 15 trillion tokens, the 8b and 30b models use 'agentic rl' to learn tool usage and code execution in sandboxed environments. alongside these, ibm released the 470m-parameter granite speech 5.0 turbo ctc, which processes 3 hours of audio in 1 second. apache 2.0 licensing of the full stack accelerates open-source development for enterprise-grade agents.