Family of code LLMs: 3B, 8B, 20B, and 34B parameters. Trained on 3.5–4.5T tokens across 116 programming languages. Supports code generation, bug fixing, code explanation, and documentation.

Granite Code 34B was competitive with CodeLlama-34B and StarCoder-15B on code benchmarks at release. Apache 2.0.

Model Details

Architecture DENSE
Parameters 34B
Training tokens 4.5T

Variants

Name Parameters Notes
Granite Code 3B 3B —
Granite Code 8B 8B —
Granite Code 20B 20B —
Granite Code 34B 34B —

Paper

Citations 10
codingopen-weight

Related