IBM unveiled Granite 4.1, an open‑source suite of decoder‑only transformers aimed at enterprise workloads. The family ships in 3 B, 8 B, and 30 B variants, all trained on a 15‑trillion‑token corpus and licensed under Apache 2.0. Unlike the prior 4.0‑H‑Small, the new 8 B model eschews mixture‑of‑experts routing, offering a dense architecture that simplifies latency budgeting. This also bundles deployment scripts for on‑prem and cloud environments, streamlining integration today.

Benchmarking shows the 8 B version outpacing its 32 B predecessor on nearly every test. On ArenaHard, it posted a 69.0 score, beating the older model’s lower result, while BFCL V3 tool‑calling rose to 68.3 versus 64.7. GSM8K math reasoning reached 92.5, and the model maintains a 512K context window without sacrificing short‑context performance. This proves meticulous data work can outweigh raw parameter growth in practice for LLMs.

IBM attributes the gains to a rigorous data pipeline that filters out hallucinations and low‑quality samples before fine‑tuning, followed by four stages of reinforcement learning. The third stage repaired math regressions introduced by earlier RLHF, while the final math‑focused run pushed scores above baseline. Customers gain lower latency across workloads. For enterprises needing predictable costs and consistent instruction following, Granite 4.1 delivers a lean, high‑performing alternative.