IBM's Granite 4.2 is a new family of dense, decoder-only reasoning LLMs released in 3B, 8B, and 30B sizes under Apache 2.0. Each model is pre-trained from scratch on ~15T tokens with a five-phase strategy extending context to 512K tokens, then fine-tuned on chain-of-thought and agentic data, and post-trained through a multi-stage GRPO-based reinforcement learning pipeline covering verifiable rewards, skill boosters, and agentic RL (SWE, terminal, search) for the 8B and 30B models. All models support thinking/non-thinking/low-effort modes and native tool calling, and are available in quantized FP8, NVFP4, MXFP4, and GGUF variants for deployment with vLLM and llama.cpp.
Table of contents
OverviewModel ArchitecturePre-TrainingSFT: Data Preparation & Quality ControlReinforcement Learning: A Multi-Stage, Multi-Environment PipelineAgentic AI Infrastructure for Scalable RLResultsQuantizationInfrastructureGetting Started (Transformers)Tool CallingMulti-Turn ConversationsUsing with Agentic Coding HarnessesQuantizationInfrastructureGetting Started (Transformers)Tool CallingMulti-Turn ConversationsUsing with Agentic Coding HarnessesQuestions this post answers
What sizes does the Granite 4.2 model family come in and what license is it released under?
Granite 4.2 ships in three dense decoder-only sizes: 3B, 8B, and 30B parameters, all released under the Apache 2.0 license. Each is pre-trained from scratch on roughly 15 trillion tokens using a five-phase strategy that extends context to 512K tokens, then fine-tuned and post-trained with multi-stage reinforcement learning. Track new open model releases like this one as they land by following IBM Granite on daily.dev.
What is the difference between the 3B, 8B, and 30B Granite 4.2 models in terms of agentic capabilities?
The 8B and 30B Granite 4.2 models go through an additional agentic reinforcement learning block covering software engineering, terminal use, and web search inside real sandboxed environments, while the 3B model stops after foundational RL and RLHF alignment without agentic training. This means only the 8B and 30B models are trained to call tools, edit and run code, and drive a terminal as agents. Compare model sizes before picking one for an agentic coding workflow on daily.dev.
How does asynchronous GRPO training work in the Granite 4.2 reinforcement learning pipeline?
Asynchronous GRPO lets generation and training run without blocking each other: generation workers keep sampling responses into a shared buffer, and once a full step's worth accumulates, the trainer takes an optimizer step and streams updated parameters back without pausing generation. A refresh can land mid-rollout, so a limit caps how many updates a worker can lag behind, and truncated importance sampling clamps the log-probability ratio to handle any staleness. Dig deeper into RL training internals like this one by following LLM training research on daily.dev.