Granite Code

Granite Code is IBM's family of decoder-only code language models, released in 3B, 8B, 20B, and 34B sizes with both base and instruction-tuned variants. The models were trained on 3 to 4 trillion tokens spanning code from 116 programming languages plus natural language, with an emphasis on ethically sourced, license-permissible data: aggressive deduplication, filtering of harmful content, ClamAV malware scanning, and PII redaction. Instruct variants add fine-tuning on code commits, math datasets, and synthetic instructions. Intended enterprise uses include code generation, explanation, bug fixing, translation between languages, and documentation, with competitive results on benchmarks such as HumanEvalPack and MBPP among open code models. Context windows range from 2K to 8K tokens depending on size. All weights are Apache 2.0 licensed for research and commercial use, downloadable from Hugging Face, with fine-tuning supported through IBM's Dolomite Engine.