IBM’s new Granite 4.2 models ride the wave of interest in local LLMs (arstechnica.com)

IBM has launched its latest Granite 4.2 models, expanding its open-weight large language model offerings to include 3B, 8B, and 30B parameter variants. These models are designed for self-hosting and feature a decoder-only architecture with an impressive 128,000-token context window. Notably, the 8B and 30B models incorporate an agentic reinforcement learning block, which enhances their abilities to perform tasks like terminal operations and web searches. The 3B variant, while still capable of using tools, does not have the same level of specialized training. What sets Granite 4.2 apart is its focus on functional reasoning, a key advancement for the AI/ML community. Rather than imitating human reasoning, these models utilize techniques like "chain-of-thought" to carry intermediate results through multiple steps, significantly improving their problem-solving capabilities. This emphasis on reasoning represents a strategic shift in model development, aligning with growing interest in local large language models that can be adapted for diverse applications, thus fostering greater innovation and experimentation within the field.