IBM has released Granite 4.2, a new open reasoning-model family built for enterprise agents, coding, tool use, and long-context workflows.
The family includes 3B, 8B, and 30B dense decoder-only language models. IBM says the models were pre-trained from scratch on roughly 15 trillion tokens, support a native 128K context length, and include long-context extension work to 512K tokens. The weights are released under the Apache 2.0 license.
The release is not only a chat-model refresh. IBM’s technical post describes a multi-stage reinforcement learning pipeline, with an agentic RL block for the 8B and 30B models that trains them to use tools inside live sandboxed environments.
Granite now targets agent work explicitly
Earlier Granite releases were enterprise-oriented instruction models. Granite 4.2 moves the center of gravity toward agents that plan, call tools, edit code, use a terminal, and search the web.
The model family has thinking, non-thinking, and low-effort thinking modes. It also supports native tool calling. IBM says that when served through an OpenAI-compatible endpoint, the model emits tool calls in the OpenAI function-calling format, which makes it easier to plug into existing agent harnesses.
The detailed training writeup matters because it shows how IBM is trying to make smaller open models useful in production settings. The RL stack separates training from rollout environments: NeMo-RL runs the training side, while NeMo-Gym hosts sandboxes, verifiers, tools, and reward calls as resources.
That architecture is relevant beyond Granite. It reflects the broader move from static instruction tuning toward live task environments, where a model learns from code execution, terminal behavior, and multi-turn tool use.
The numbers are useful but provider-authored
IBM reports benchmark results across agentic coding, general tool use, reasoning, chat, instruction following, and long context. The 30B model reports 57.00 on SWE-Bench Verified, 29.24 on Terminal-Bench 2.1, 89.17 on AIME25, and 81.38 on RULER 128K.
Those results should be treated as release evidence, not as a complete independent ranking. The model leaderboard that matters for buyers will depend on third-party evals, serving providers, quantized performance, latency, and whether the model holds up in the specific agent harness a team uses.
IBM also released quantized variants for inference. The Hugging Face post describes FP8, NVFP4, MXFP4, and GGUF paths, with vLLM, SGLang, llama.cpp, and common local-runtime routes represented across the release materials.
That makes Granite 4.2 more practical than a single full-precision checkpoint. The open-model story is not just whether weights are downloadable. It is whether teams can run the model cheaply, govern it internally, fine-tune it, and connect it to real tools without giving up control of the stack.
