IBM

Granite Code

IBM's enterprise-focused code models from 3B to 34B, trained on 116 programming languages with license-permissible data and released under Apache 2.0.

Hugging Face GitHub Ollama

Overview

Granite Code is IBM's family of decoder-only code models, part of the broader Granite series used in watsonx. IBM emphasized data governance: the training corpus was filtered for permissive licenses and scrubbed of PII. Long-context 128K versions of the 3B and 8B models were released later.

Strengths

Clean licensing story for regulated industries
Solid code generation, explanation and fixing across 116 languages
Range of sizes for edge and server deployment

Run it locally

ollama run granite-code:8b

Good fit for

Enterprises that need legal clarity on training data along with an Apache 2.0 license.

Frequently asked questions

What makes Granite Code enterprise-friendly?

IBM trained it only on license-permissible data, published details of its data pipeline and released the weights under Apache 2.0.

Similar models

all models →

~/models/qwen2-5-coder

Alibaba Qwen

Qwen2.5-Coder

A dense code model family from 0.5B to 32B parameters. The 32B Instruct model reached GPT-4o-level code generation at release and is easy to run locally.

params: 32B
context: 128K
released: 2024-11

Commercial OKApache 2.0

~/models/codestral

Mistral AI

Codestral

Mistral AI's 22B open-weight code model, trained on 80+ programming languages, with strong fill-in-the-middle support for IDE autocomplete. Non-production license.

params: 22B
context: 32K
released: 2024-05

RestrictedMistral AI Non-Production License

~/models/starcoder2

BigCode (Hugging Face, ServiceNow, NVIDIA)

StarCoder2

Fully open code models (3B, 7B, 15B) trained on The Stack v2 with transparent data, published training details and opt-out support for code authors.

params: 15B
context: 16K
released: 2024-02

Commercial OKBigCode OpenRAIL-M