IBM releases Granite 4.1-8B with 131K context window and enhanced tool-calling capabilities

IBM has released Granite 4.1-8B, an 8-billion parameter instruction-following model with a 131,072-token context window. The model was released on April 29, 2025, under an Apache 2.0 license.

Performance benchmarks

According to IBM, Granite 4.1-8B achieves the following scores:

• Code tasks: 85.37% on HumanEval pass@1, 87.30% on MBPP pass@1, 79.88% on HumanEval+ pass@1
• General tasks: 73.84% on MMLU 5-shot, 80.51% on BBH 3-shot with chain-of-thought
• Math tasks: 92.49% on GSM8K 8-shot, 80.10% on Minerva Math 0-shot with CoT
• Tool-calling: 68.27% on BFCL v3
• Alignment: 87.06% on IFEval average

The model is part of a three-model family including 3B and 30B parameter versions.

Technical specifications

Granite 4.1-8B uses a decoder-only dense transformer architecture with:

• 4,096 embedding size
• 40 layers
• 32 attention heads with 8 key-value heads
• Grouped Query Attention (GQA)
• RoPE positional embeddings
• SwiGLU activation in MLP layers
• 12,800 MLP hidden size

Training and capabilities

IBM trained the model on a combination of open source instruction datasets with permissive licenses and internally generated synthetic data. The post-training pipeline included supervised fine-tuning and reinforcement learning alignment.

The model supports 12 languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. According to IBM, it can be fine-tuned for additional languages.

Key capabilities include:

• Text summarization and classification
• Question-answering and RAG
• Code generation and completion
• Function calling with OpenAI-compatible tool definitions
• Fill-in-the-middle code completions
• Multilingual dialog

The model achieves 64.84% on MMMLU 5-shot across 11 languages and 58.89% on INCLUDE 5-shot across 14 languages.

Safety benchmarks

IBM reports safety scores of 95.80% on SALAD-Bench and 81.19% on AttaQ for the 8B model.

Availability

The model is available on Hugging Face under the Apache 2.0 license. Pricing for API access has not been disclosed. IBM provides code examples for both basic text generation and tool-calling use cases using the Transformers library.

What this means

Granite 4.1-8B represents IBM's push into the competitive 8B parameter model space with strong code performance and a notably large 131K context window. The Apache 2.0 license and multilingual support position it as an alternative to models like Llama 3.1 8B and Mistral 7B for enterprises requiring permissive licensing. The tool-calling improvements and comprehensive benchmark suite suggest IBM is targeting production AI assistant deployments, though actual inference costs and API availability remain unclear.