AI
IBM Releases Granite 4.2: Dense reasoning models built on pure Transformer architecture
IBM's new Granite 4.2 family prioritizes reasoning capabilities and efficient agentic reasoning across 3B, 8B, and 30B parameter models while maintaining its all-attention, dense architecture.
The New Stack
IBM unveiled its latest generation of open-weight language models this week, doubling down on dense Transformer-based designs as many competitors pivot toward hybrid architectures. The Granite 4.2 family arrives in three sizes—3 billion, 8 billion, and 30 billion parameters—with an explicit focus on reasoning-driven workloads.
An unconventional architectural choice#
The move distinguishes IBM from rivals experimenting with Mamba/attention hybrids, such as Nvidia's Nemotron 3 family. While IBM tested hybrid and mixed-of-experts variants in the Granite 4.0 generation, the team retreated to conventional all-attention, dense Transformers with version 4.1—and has retained that approach for 4.2. IBM's rationale emphasizes simplicity and flexibility for downstream fine-tuning tasks.
Reasoning as a core feature#
The 4.2 release marks reasoning as a central capability. Models operate in both thinking and non-thinking modes, with an efficiency-focused variant that uses minimal reasoning tokens for straightforward queries. This represents a shift from IBM's earlier position that reasoning models were too costly for enterprise adoption; the company now views reasoning as a necessary—though optional—feature rather than a luxury.
Unlike competitors such as Qwen-3.8 27B and Google Gemma 4, Granite 4.2 remains text-only, though IBM continues to offer multimodal options through products like Granite Vision 4.1 4B.
Training at scale with synthetic data#
All three models underwent pre-training on 15 trillion tokens across five phases, culminating in support for a 512,000-token context window (with 128K native support). A notable component: the training corpus includes 1 trillion tokens of synthetically generated code via IBM's CodeAlchemy pipeline, though the resulting coding performance remains middling.
The 8B and 30B variants received additional agentic reinforcement learning to handle tool calling, code execution, terminal access, and web search. Combined with human feedback alignment, this training approach readies these models for multi-step, complex agent workflows. The 3B model also supports tool calling, though with more limited capability.
