ModelsFEATURED

Granite 4.1 LLMs: How They’re Built

IBM's Granite 4.1 open-source LLM family (3B–30B params, 512K context) achieves superior performance to its larger 32B MoE predecessor through dense architecture and multi-stage training on 4.1M curated samples with GRPO reinforcement learning.

Thursday, April 30, 2026 12:00 PM UTC2 MIN READSOURCE: Hugging FaceBY sys://pipeline

IBM releases Granite 4.1, a family of open-source LLMs (3B, 8B, 30B parameters) under Apache 2.0, trained on 15 trillion tokens with multi-stage data curation and advanced training techniques. The models extend to 512K context length and undergo supervised fine-tuning on 4.1M curated samples plus multi-stage reinforcement learning using GRPO with DAPO loss. The 8B instruct variant matches or exceeds performance of IBM's previous 32B MoE model despite its simpler, more efficient dense architecture.

Read original at Hugging Face

The IBM Granite 4.1 family of models

IBM Granite 4.1 trades expensive reasoning for tool-calling efficiency in its 3B–30B model lineup, competing favorably with Gemma and Qwen on practical enterprise tasks.