AI Potluck
Model components / Fine-tuned / chat models

Granite Code Instruct

IBM

IBM's enterprise code model line, fine-tuned from Granite Code Base for code instruction-following at 3B, 8B, 20B and 34B with 2K and 128K context variants. It was trained on permissively licensed code, math and natural-language instruction data, which is IBM's stated basis for offering enterprises clarity on training-data provenance.

The model card now opens with a deprecation warning; the line has been folded into the unified Granite 4.x mainline. IBM names its data sources but does not release the processed datasets. Verified 2026-08-13 via the model card, the granite-code-models repository and IBM's release blog.

Openness

3 high confidence
3.0
weights
open(Apache-2.0,on HF)
code
partial(inference + fine-tune sample scripts via Dolomite Engine
data
described_not_released(CommitPackFT/MathInstruct/Glaive/HelpSteer/Open-Platypus named in card but processed corpus not redistributable)
post-training-data
described(SFT mixture named, not released)
paper
open(arXiv 2405.04324)
model_card
open
license
Apache-2.0(OSI)

An OSI Apache-2.0 license with downloadable weights, a paper and a model card, but the training and post-training data are only described rather than released (the processed corpus is not redistributable) and no full training pipeline ships, so this is open weights rather than open source. It is strong within the open-weights class: genuinely permissive, with no use caps. The card now opens with a deprecation banner advising against use in new projects, which does not move openness.

  • https://huggingface.co/ibm-granite/granite-3b-code-instruct-2k recorded 2026-08-13

    `license:apache-2.0` in the repo tags and `"gated":false`; training sets NAMED as dataset tags (bigcode/commitpackft, TIGER-Lab/MathInstruct, meta-math/MetaMathQA, glaiveai/glaive-code-assistant-v3, garage-bAInd/Open-Platypus, nvidia/HelpSteer) with no processed mixture published; arxiv:2405.04324; model-index with HumanEvalSynthesis scores; a DEPRECATION WARNING banner at the top of the card

  • https://github.com/ibm-granite/granite-code-models recorded 2026-08-13

    repo with Apache-2.0, fine-tune sample scripts (Dolomite Engine), but no training data download and no full training pipeline

  • https://research.ibm.com/blog/granite-code-models-open-source recorded 2026-08-13

    IBM describes data sources (GitHub Code Clean, StarCoderData, CodeNet) but does not release the processed datasets

Adoption

1 medium confidence
1.0

Officially deprecated and superseded by Granite 4.x, with little active pull volume left. The declared artifact, granite-34b-code-instruct-8k, reads 628 downloads in the trailing 30 days and the 3b variant reads 5,601; both are under 10K, as is any sum across the family's Instruct sizes. Hobbyist and legacy-research scale.

Capability

2 high confidence
2.0

Strong-for-size 2024 code SLMs (matched open code LLMs of their era), but sub-13B and ~51% HumanEval Python is far below 2026 frontier chat/agentic coders (Opus 4.8 88.6 SWE-bench). Low-mid on a frontier-relative scale.

Verified 2026-08-13