Granite 4.0

weights

open(Apache-2.0, ungated on HF across the 4.0 line)

data

closed(SFT data characterized in the card (permissive public + internal synthetic + human-curated) but neither the pretraining corpus nor the mixtures are released)

code

partial(ibm-granite/granite-4.0-language-models is docs/usage, no pretraining pipeline)

license

Apache-2.0(OSI)

Apache-2.0 weights, ungated in every 4.0 SKU read, with training data described in categories rather than released and no training pipeline. That is open weights with a closed recipe, 3 by the fallthrough. The card's data description ("publicly available datasets with permissive license, internal synthetic data, human-curated data") characterizes sources without fully specifying composition, so it does not reach documented-not-released, and the outcome is 3 either way because the documented-not-released rung also requires open code.

• https://huggingface.co/api/models/ibm-granite/granite-4.0-h-small recorded 2026-09-01

  `gated: false`, cardData.license `apache-2.0`, safetensors listed; 25,719 downloads in the trailing 30 days for the flagship h-small SKU.

• https://huggingface.co/ibm-granite/granite-4.0-h-small/raw/main/README.md recorded 2026-09-01

  Card Training Data section: "our SFT data is largely comprised of three key sources: (1) publicly available datasets with permissive license, (2) internal synthetic data targeting specific capabilities, and (3) a select set of human-curated data" — described in categories, not released. Infrastructure section names the GB200 cluster; no training pipeline links.

• https://api.github.com/repos/ibm-granite/granite-4.0-language-models recorded 2026-09-01

  `license: Apache-2.0`, 210 stars; the 4.0 release repository, which carries no pretraining pipeline — what holds `code` at partial.

194,509 downloads in the trailing 30 days across the four declared artifacts (granite-4.0-h-tiny 88,627; granite-4.0-micro 60,053; granite-4.0-h-small 25,719; granite-4.0-h-micro 20,110), which bands at level 3 (100K-1M) on the model adoption scale. The smaller on-device SKUs carry most of the reach.

• https://huggingface.co/api/models/ibm-granite/granite-4.0-h-tiny recorded 2026-09-01

  88,627 downloads in the trailing 30 days for ibm-granite/granite-4.0-h-tiny

• https://huggingface.co/api/models/ibm-granite/granite-4.0-micro recorded 2026-09-01

  60,053 downloads in the trailing 30 days for ibm-granite/granite-4.0-micro

• https://huggingface.co/api/models/ibm-granite/granite-4.0-h-small recorded 2026-09-01

  25,719 downloads in the trailing 30 days for ibm-granite/granite-4.0-h-small

• https://huggingface.co/api/models/ibm-granite/granite-4.0-h-micro recorded 2026-09-01

  20,110 downloads in the trailing 30 days for ibm-granite/granite-4.0-h-micro

Mid-tier: a capable efficiency-focused line, competitive within the small-to-mid open-weight class and well short of the open-weight frontier the category's 4s and 5s hold (olmo, deepseek, qwen, kimi). Sits level with llama, whose own reading is taken on Llama 4 Maverick — the same non-reasoning open-weight mid-field.

• https://huggingface.co/ibm-granite/granite-4.0-h-small/raw/main/README.md recorded 2026-09-01

  The h-small card's evaluation tables benchmark the 4.0 line against similar-size open-weight peers on instruction-following and enterprise task suites, positioning it as competitive in its size class rather than frontier.