IBM releases Granite 4.1 dense LLMs with Apache 2.0 license
ibm released granite 4.1, a family of dense decoder-only llms (3b, 8b, 30b) trained on 15t tokens. the 8b instruct model matches the previous 32b moe performance. features include 512k context windows, multi-stage annealing, and apache 2.0 licensing.
IBM released the Granite 4.1 family, a set of dense, decoder-only LLMs available in 3B, 8B, and 30B parameters. The models are trained from scratch on approximately 15 trillion tokens using a five-phase pipeline concluding with long-context extension to 512K tokens. Despite using a simpler dense architecture with fewer parameters, the 8B Instruct model matches or exceeds the performance of the previous generation's 32B Mixture-of-Experts model.
Technical specifications include Grouped Query Attention (GQA), RoPE, and SwiGLU activations. Post-training involves supervised fine-tuning on 4.1M curated samples and reinforcement learning via GRPO with DAPO loss. The entire model family is released under the Apache 2.0 license.