Piloting the world's first double-blind AI evaluations
deepmind.google·1d ago
AI Engineering
Hugging Face Blog·3d ago·5 min read
Granite 4.2 LLMs: How They're Built
TL;DR
Granite 4.2 introduces a family of dense, decoder-only reasoning language models in sizes 3B, 8B, and 30B, enhancing reasoning capabilities with a five-phase training strategy. Each model supports native tool calling and operates in different thinking modes for varied task complexity.
✦ Why It Matters
Explore Granite 4.2 for advanced reasoning capabilities in your AI applications today.
Key Takeaways
- Experiment with different model sizes for specific tasks.
- Utilize the thinking/non-thinking switch for efficiency.
- Incorporate tool calling in your applications.
- Leverage the extensive training data for improved performance.
How It Works
Granite 4.2 employs a dense transformer architecture with components like Grouped Query Attention and Rotary Position Embedding. The five-phase training strategy includes foundational pre-training, supervised fine-tuning, and a multi-stage reinforcement learning pipeline that enhances reasoning and tool usage capabilities.
The models are designed to operate in various modes, allowing them to adapt their reasoning efforts based on task complexity.