Granite 4.2 30B

Granite 4.2 30B is an open-weight, dense decoder-only transformer model developed by IBM as the flagship of the Granite 4.2 reasoning family. Released under the Apache 2.0 license, it is designed for complex enterprise tasks that require multi-step logic, code generation, and agentic workflows. Unlike previous hybrid architectures, the 4.2 generation utilizes a standard dense transformer design to ensure broad compatibility across diverse hardware and serving environments.

A defining feature of the model is its native support for switchable thinking modes. Users can toggle between three distinct operational states: a default thinking mode that uses visible chain-of-thought reasoning within <think> tags, a non-thinking mode for direct answers with lower latency, and a low-effort mode for simpler queries. This flexibility allows the model to balance depth of deliberation against computational cost on a per-query basis.

Agentic Capabilities and Training

The model is specifically optimized for agentic tool-calling and autonomous software engineering tasks. During its post-training phase, it underwent multi-stage reinforcement learning (RL) within live sandboxed environments, including terminal shells and web-search interfaces. This training enables the model to reason about which tools to invoke and why, significantly improving performance on benchmarks such as SWE-bench Verified and BFCL (Berkeley Function Calling Leaderboard).

Granite 4.2 30B supports a native context window of 128,000 tokens, with extensions up to 512,000 tokens. It is trained on approximately 15 trillion tokens, including a substantial portion of synthetic code and reasoning-trajectory data. The model supports multilingual dialogue across 12 languages, including English, German, Japanese, and Arabic.