IBM has released Granite 4.2, the latest generation of its open-weight large language models designed to be downloaded and run locally rather than accessed via the cloud. The launch taps into growing demand for self-hosted AI, as developers and enterprises look for cheaper alternatives to costly frontier models from the likes of OpenAI and Anthropic, and it reflects IBM's continued focus on dependable, predictable deployments rather than chasing headline-grabbing performance.

The new family comes in 3B, 8B and 30B parameter sizes, all using a decoder-only architecture with a native 128,000-token context window. The 8B and 30B models have undergone additional agentic reinforcement learning, equipping them to use terminals, search the web and call external tools, while the smaller 3B model supports tools but lacks this specialised training. IBM describes Granite 4.2 as its "reasoning-focused" release, meaning the models are built to work through problems step by step via chain-of-thought techniques, improving accuracy in some cases at the cost of slower responses and higher compute demands.