IBM has launched its Granite 4.2 open-weight LLM family, featuring 3B, 8B, and 30B parameter variants with a 128K token context window. This release is notably reasoning-focused, emphasizing chain-of-thought capabilities, and the 8B/30B variants include agentic reinforcement learning for tasks like terminal use, web search, and tool utilization. The models target self-hosted enterprise deployments, prioritizing reliability over cutting-edge speed.
Background
IBM's Granite series targets enterprise developers seeking open-weight, self-hostable LLMs as an alternative to proprietary cloud-based models. Local LLM adoption has grown significantly amid rising costs and compute constraints of frontier cloud models.
- Source
- Ars Technica
- Published
- Aug 26, 2026 at 07:10 PM
- Score
- 6.0 / 10