IBM Granite Code Models by IBM Research
Last refreshed 2026-10-02. Next refresh: weekly.
Details
About
Granite Code is IBM Research's family of open decoder-only models for code tasks such as generating, fixing, explaining, and documenting code. They were trained on code in 116 programming languages, using data IBM describes as license-permissible and collected under its AI ethics principles. The family comes in base and instruct versions at 3B, 8B, 20B, and 34B parameters; the instruct models are fine-tuned on Git commits paired with human instructions plus open code and math instruction data. IBM now marks these models as not recommended for new projects and points users to its mainline Granite language models, whose code abilities supersede them.
Decision facts
- Best fit
- codingmath-heavy prompts
- Capability starting point
- Granite Code
- Lowest tracked input
- Not tracked
- Closest related family
- Granite 4
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| Granite Code | Use when provider availability and model metadata match the workload. | 2024-05 | — | Current |
Release Timeline
1 release groupSpecifications(1 models)
| Model | Released |
|---|---|
| Granite Code | 2024-05 |
