IBM just dropped its biggest Granite release yet. The 4.1 family includes language models in 3B, 8B, and 30B parameter sizes, plus new versions of Granite Vision, Speech, Guardian, and Embedding models. All Apache 2.0 licensed.
But the real story is architectural. IBM moved away from Mixture-of-Experts designs to dense, decoder-only models. The claim is bold: the new 8B instruct model reportedly matches or beats the older 32B MoE model according to IBM's performance reports. That's a 4x efficiency improvement if it holds up in practice. The reasoning is straightforward. MoE models introduce latency variability in production because routing between expert sub-networks isn't predictable. Dense models give you stable inference. According to Rameswar Panda, a distinguished engineer at IBM Research and key architect of the models, "Granite 4.1 delivers competitive instruction-following and tool-calling performance without relying on long chains of thought, offering predictable latency, stable token usage, and lower operational cost." For enterprises running agents at scale, that predictability matters more than chasing benchmark scores.
The broader suite covers real production workflows like document processing and agent routing. Granite Vision 4.1 handles document understanding, extracting structured content like tables and key-value pairs from business documents. Granite Speech 4.1 hits a 5.33% word-error rate on the OpenASR Leaderboard. Granite Guardian handles harm detection. The language models support context windows up to 512K tokens, trained on roughly 15 trillion tokens with staged refinement and multi-stage reinforcement learning.
Everything ships on Hugging Face, Replicate, watsonx, Ollama, and other platforms. IBM built this for production where cost and speed beat raw intelligence. A pragmatic bet that fits what most agent developers actually need.