Granite 4.2 brings native reasoning to enterprise agents
IBM Research ● Covered by 5 sources
IBM launched Granite 4.2 with built-in reasoning for enterprise agents. It’s meant to plan and use tools before acting, not just spit out answers.
Based on reporting by IBM Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
IBM’s Granite family just got a more ambitious update. The new Granite 4.2 language models come in 3B, 8B, and 30B sizes, and the pitch is clear: these aren’t chatty text generators, they’re meant for agents that have to think, call tools, and finish messy work.
The big shift is “thinking” built into the model itself. IBM says Granite 4.2 can step through problems, weigh trade-offs, and catch mistakes before they turn into bad actions. That matters most in enterprise workflows, where a task might mean finding the right information, choosing the right application, using it in the right order, and checking the result afterward.
The training setup was redesigned to get there. IBM built on Granite 4.0, then added a multi-stage reinforcement learning process. All three models went through a foundational RL stage for math, science, coding, reasoning, and tool use. The 8B and 30B versions then got another pass aimed at enterprise-style work such as software engineering, terminal-based coding, and search-driven workflows, alongside RLHF alignment.
There are a few other pieces pushing the models forward. IBM trained them on 1 trillion tokens of synthetic code from its CodeAlchemy pipeline, added a mid-training step it says improves reasoning, and used speculative decoding so they can generate text faster and serve more users. The models are also under Apache 2.0, so organizations can download, fine-tune, and ship them without licensing restrictions.
IBM is also leaning hard into deployment flexibility. Granite 4.2 is built for cloud, on-premises, and edge use, and smaller sizes are meant for high-throughput work while larger ones handle deeper reasoning and more complex coding. On the speech side, IBM introduced Granite Speech 5.0 Turbo CTC and 5.0 Turbo CTC NC, each with 470 million parameters and no LLM backbone, aimed at fast transcription on laptops, smartphones, and other edge devices. In testing, IBM says one of them reached about 12,600 RTFx on a single H100 GPU, versus around 6,000 for current speed leaders on Hugging Face’s Open ASR leaderboard.
The positioning is practical rather than flashy. IBM is trying to make agentic AI look less like a demo and more like infrastructure: faster, smaller, and easier to actually deploy without a licensing fight.
My take — AI-written commentary, not fact-checked reporting
This is the rare enterprise AI pitch that sounds like someone has met a real workflow before lunch. Open weights plus Apache 2.0 matters more than another grand speech about agents, because companies keep asking for control and then acting shocked when control has a bill attached. IBM is betting on useful, boring reliability, which is usually where the money ends up.
Read more about this at: IBM Research