IBM Releases Granite 4.1 30B With 131K Context Window and Enhanced Tool-Calling

IBM released Granite 4.1 30B, a 30-billion parameter instruction-following model with a 131,072 token context window. Released April 29, 2026 under Apache 2.0 license, the model is available on Hugging Face.

Benchmark Performance

The model scores 80.16 on MMLU 5-shot, 88.41 on HumanEval pass@1, and 85.45 on MBPP pass@1. On reasoning benchmarks, it achieves 83.74 on BBH 3-shot with chain-of-thought and 64.09 on MMLU-Pro 5-shot with CoT. The model scores 89.65 on IFEval (instruction following) and 71.02 on ArenaHard.

For math tasks, Granite 4.1 30B reaches 94.16 on GSM8K 8-shot and 81.93 on DeepMind Math 0-shot with CoT. Tool-calling capability scores 73.68 on BFCL v3.

Architecture and Training

The model uses a decoder-only dense transformer with grouped-query attention, RoPE positional embeddings, and SwiGLU activation. It has 64 layers, 32 attention heads (8 KV heads), 4,096 embedding size, and 32,768 MLP hidden size. Attention head size is 128.

IBM trained the model using "a combination of open source instruction datasets with permissive license and internally collected synthetic datasets," according to the model card. The post-training pipeline included supervised fine-tuning and reinforcement learning alignment.

Language and Safety

Granite 4.1 30B officially supports 12 languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. It scores 73.71 on MMMLU 5-shot (11 languages) and 67.26 on INCLUDE 5-shot (14 languages).

On safety benchmarks, the model achieves 96.41 on SALAD-Bench, 85.76 on AttaQ, and 78.19 average on Tulu3 Safety Eval.

Tool-Calling Implementation

The model supports tool-calling using OpenAI's function definition schema. IBM's implementation uses XML tags (<tool_call>) to structure function calls with JSON objects containing function names and arguments. The model can integrate with external APIs and functions.

What This Means

Granite 4.1 30B provides open-source teams an Apache 2.0 licensed alternative with competitive performance on instruction-following and code generation tasks. The 131K context window and multilingual support position it for enterprise RAG applications. However, pricing for hosted inference is not yet disclosed, and the model's performance trails frontier models on advanced reasoning benchmarks like GPQA (45.76 vs. 50+ for leading models). The enhanced tool-calling capability and permissive license make it particularly relevant for commercial deployments requiring function integration.