IBM’s new Granite 4.2 models add reasoning and stay dense
On Tuesday, IBM launched the latest family of its open-weight Granite large language models (LLMs). Weighing in at 3 billion, The post IBM’s new Granite 4.2 models add reasoning and stay dense appeared first on The New Stack .
IBM unveiled its latest family of open-weight Granite large language models (LLMs) on Tuesday, featuring 3 billion, 8 billion, and 30 billion parameters. The company is taking a unique approach compared to its competitors, focusing on dense, decoder-only reasoning models that are pre-trained from scratch. While many recent models have shifted towards hybrid Mamba/attention architectures, IBM returned to the all-attention, dense Transformer architecture with Granite 4.1 before releasing the reasoning-focused Granite 4.2 family.
The 4.2 models support both reasoning and non-thinking modes, with an additional low-effort mode that minimizes reasoning for easy questions. Although not multi-modal like some competitors, IBM offers Granite Vision 4.1 4B for text-image tasks. Pre-training involved 15 trillion tokens, including 1 trillion synthetic code tokens, and IBM incorporated reinforcement learning from human feedback (RLHF) to improve complex, multi-step agentic tasks.
In benchmark tests, the smaller 8B model often performs comparably to the larger 30B model, making it accessible even on lower-end GPUs. IBM emphasizes that the main advantage of the Granite 4.2 models lies in their performance in high-throughput agentic tasks, enabling AI systems to plan, call applications, and execute complex tasks reliably while remaining cost-effective for enterprises.
Written by urgent.news from The New Stack's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
