granite-4.2

Model Summary

Model Overview

What is IBM Granite?

Granite is a family of open-source large language models developed by IBM, designed for enterprise and research use. Granite models are built to be versatile, safe, and efficient — covering a range of sizes and capabilities from compact edge-deployable models to large-scale reasoning systems. All Granite models are released under the Apache 2.0 license, enabling unrestricted commercial and academic use.

The Granite 4.2 generation introduces native reasoning (thinking) capabilities, allowing models to perform step-by-step chain-of-thought reasoning before producing final answers. This significantly improves performance on complex math, coding, multi-step logic, and agentic tool-calling tasks.

Description

Granite-4.2-8B is the mid-size reasoning model in the Granite 4.2 family. It delivers strong performance on reasoning-intensive tasks by leveraging built-in ... chain-of-thought. It supports flexible thinking modes — full thinking (default), non-thinking, and low-effort — allowing users to balance depth vs. latency on a per-query basis.

Key capabilities:

• Built-in Reasoning: Native chain-of-thought that significantly improves performance on math, coding, and complex multi-step problems.
• Flexible Thinking Modes: Seamlessly switch between full thinking, non-thinking, and low-effort modes within a single model.
• Reasoning-Augmented Tool Calling: The model reasons about which tools to invoke and why, producing more accurate function calls.
• 512K Context Window: Supports long documents, multi-turn conversations, and complex agentic workflows.
• Apache 2.0 Licensed: Fully open for commercial and research use.

Model Design

Granite-4.2-8B is built on a decoder-only dense transformer architecture with the following core components:

• Attention: Grouped Query Attention (GQA) with 32 attention heads and 8 KV heads
• Position Embedding: Rotary Position Embedding (RoPE) with θ = 10,000,000
• Feed-Forward: MLP with SwiGLU activation (hidden size 12800)
• Normalization: RMSNorm (ε = 1e-5)
• Embeddings: Separate input/output embeddings (not tied)
• Precision: bfloat16

Training Methodology

Granite-4.2-8B is post-trained from Granite-4.1-8B-Base through a rigorous multi-stage pipeline that progressively unlocks reasoning, tool use, and instruction-following capabilities. A full listing of training datasets is available in the Granite 4.2 GitHub repository. The training pipeline consists of three stages:

Stage 1: Pre-Training

Granite-4.2-8B builds on Granite-4.1-8B-Base, which was pre-trained on a large-scale English as well as multilingual corpus. For full pre-training details (data composition, training recipe, and infrastructure), refer to our Granite 4.1 Technical Blog.

Stage 2: Supervised Fine-Tuning

The SFT stage draws on instruction-following, chain-of-thought, and reasoning data to cultivate the model's reasoning and thinking abilities. For all the three, 3B, 8B and 30B models, the training corpus comprises four sources: (1) publicly available datasets under permissive licenses, (2) internally generated synthetic data targeting reasoning, tool calling, and chain-of-thought capabilities, (3) agentic traces collected across a diverse range of tasks, and (4) a curated selection of human-authored data. Hyperparameters were tuned before training was scaled to all three model sizes. For the 30B model, we conducted a second SFT phase, in which the agentic data was up-sampled while a smaller share of general replay data was retained. This phase trained for a single epoch, starting from a lower learning rate than Phase 1.

Stage 3: Reinforcement Learning

The final stage of training applies multi-phase, multi-environment reinforcement learning using Group Relative Policy Optimization (GRPO). Training spans a broad mix of environments including math, code, science, instruction following, tool use, general chat and structured output. Most environments provide verifiable rewards, while open-ended prompts are scored by a generative reward model. Training runs asynchronously: generation and policy updates occupy separate GPU pools rather than proceeding in lockstep, and weights are refreshed in flight.

After the reward-driven phases, a preference-alignment (RLHF) phase tunes helpfulness, conversational quality, and safety. Reinforcement learning is carried out with NeMo RL, and the RL environments run on Ne