Granite 4.2 Brings Open Reasoning to 30B Models

IBM Research and the IBM Granite Team released Granite 4.2 on August 25, 2026, as an open family of reasoning-focused language models. The 3B, 8B, and 30B dense decoder-only models add native step-by-step reasoning, switchable thinking modes, and reinforcement learning in software environments under the Apache 2.0 license.

For developers, the release combines a long context window with training aimed at coding, terminal use, web search, and multi-step task completion. Granite 4.2 is available across the open model ecosystem, including Transformers, vLLM, Ollama, LM Studio, and official GGUF quantizations.

Model Sizes and Architecture

Every Granite 4.2 model uses the GraniteForCausalLM dense decoder-only transformer architecture. Grouped Query Attention uses eight KV heads, while the attention configuration includes 40 attention heads for the smaller models and 32 for the 30B model.

Component	Granite 4.2 specification
Model sizes	3B, 8B, 30B parameters
Context	128K native pre-training, extended to 512K tokens
Attention	Grouped Query Attention, 8 KV heads
Position encoding	RoPE, θ = 10,000,000
Feed-forward network	SwiGLU MLP
Normalization	RMSNorm, ε = 1e-5
Embeddings	Separate input and output embeddings, trained in bfloat16
License	Apache 2.0

The 512K-token extension is particularly relevant to repository analysis, large document processing, and long-running agent sessions. Context capacity still carries memory and latency costs, especially when the model is generating extended reasoning, so applications should reserve the full window for tasks that need it.

Reasoning Modes and Tool Calling

Switchable reasoning is built into each model through <think>...</think> blocks. A request can use full thinking for complex mathematics, coding, or multi-step logic; non-thinking mode for lower-latency responses; or low-effort mode for simpler tasks with a restricted reasoning budget.

This per-query control gives application developers a direct quality-latency tradeoff. A support classifier or routing step can use non-thinking mode, while a coding agent can enable full thinking only after a task crosses a complexity threshold. The design fits established chain-of-thought prompting patterns, but the mode is part of the model’s intended operating behavior rather than an application-only prompt convention.

Granite 4.2 also includes a standardized tool-calling interface compatible with OpenAI endpoints, including deployments through vLLM. Developers building AI agent frameworks can therefore connect the models to existing tool routers without creating a model-specific calling format.

Training for Verifiable Work

Pre-training used roughly 15 trillion tokens across five phases. The first two phases established general capabilities, phases three and four annealed higher-quality data, and the fifth extended the model to a 512K-token context.

Supervised fine-tuning used approximately 7.2 million samples covering reasoning, instruction following, synthetic code and mathematics, human-authored examples, and multi-turn agent trajectories. The data included IBM’s open-source CodeAlchemy dataset. The 30B model received a second SFT phase with additional agentic-data upsampling.

Post-training used asynchronous Group Relative Policy Optimization, which compares responses against a group-mean reward baseline without a separate value network. Foundational reinforcement learning covered all three sizes with verifiable rewards from math checkers, unit tests, and format validation.

The 8B and 30B models also received agentic RL in sandboxed environments. SWE Agent worked on real software repositories through the OpenHands harness and hidden tests. Terminal Agent handled interactive shell sessions for up to 64 environment turns through Harbor and Terminus-2. Search Agent navigated the live web through multi-hop queries evaluated by language-model judges.

A final preference and safety stage applied a reasoning-length penalty, pushing the models away from unnecessarily verbose internal work. If you deploy reasoning models in production, evaluating AI agents should include both task success and the token cost of reaching that result.

Infrastructure and Deployment

Training ran on an NVIDIA GB200 NVL72 cluster hosted by CoreWeave, with a 72-GPU NVLink domain and 400 Gb/s InfiniBand fabric. NeMo-RL managed training through Megatron-Core, vLLM rollout generation, and Megatron-Bridge weight conversion, while NeMo-Gym handled environments, tools, sandboxes, and reward evaluation.

For deployment, start with the smallest model that satisfies the task. The 3B model suits low-resource inference and routing, the 8B model is the entry point for agentic RL capabilities, and the 30B model provides the largest reasoning capacity in the family. Benchmark your own workload across thinking modes, context length, tool-call reliability, and end-to-end token cost before selecting a production default.