Model release

Granite 4.2 3B

Granite 4.2 3B is an open-weight model release in the IBM Granite family from IBM (Granite Team)1, version 4.2 (3B), released Aug 25, 2026. Its weights can be downloaded by the public1 under the Apache License 2.01.

Granite 4.2 3B is a 3-billion-parameter dense decoder-only language model from IBM with built- in reasoning, selectable thinking, non-thinking, and low-effort modes, and tool calling. The model card states native 128K-token context support with long-context extension to 512K, while the configuration file shipped with the weights sets a maximum of 131,072 positions.1 2Fact reviewed Oct 2, 2026

Model hub: Model card (external site: huggingface.co)
Repository: Granite 4.2 language models repository (external site: github.com)
Release notes: Technical blog (external site: huggingface.co)
License: Apache License 2.0 (repository) (external site: github.com)

On AI on Hugging Face: ibm-granite/granite-4.2-3b

Last reviewedSep 29, 2026Entry updated Oct 2, 2026Documented release Aug 25, 2026

Availability and license

Overall availability

Public

Weights are downloadable from Hugging Face without a gating agreement; Apache 2.0 terms apply.1

Availability fact review: not separately recorded

Availability is separate from permission: read the license before using or redistributing.

Apache License 2.0 (external site: apache.org)1
Applies to: Weights · SPDX Apache-2.0 · on the rubric’s OSI-approved software-license list · fact-level rights review not recorded
Plain-language guide to the Apache License 2.0
Apache License 2.0 (external site: github.com)3
Applies to: Code · SPDX Apache-2.0 · on the rubric’s OSI-approved software-license list · fact-level rights review not recorded
Plain-language guide to the Apache License 2.0
Community Data License Agreement – Permissive, Version 2.0 (external site: github.com)4
Applies to: Documentation · SPDX CDLA-Permissive-2.0 · fact-level rights review not recorded
Plain-language guide to the Community Data License Agreement – Permissive, Version 2.0

The Apache 2.0 license is stated in the model card metadata and summary table. IBM's machine-readable training-data disclosures are published separately under CDLA Permissive v2.1 4

Component reuse rights

A readable or downloadable component is not automatically reusable. These indicators concern recorded license evidence, not system certification.

weights: Unknown — no complete fact-level rights review
code: Unknown — no complete fact-level rights review
data: Unknown — no complete fact-level rights review
documentation: Unknown — no complete fact-level rights review

No complete system-rights review is recorded for this release.

Model-disclosure tier

Computed from the checklist below using USASI rubric v0.2. An editorial category, not a certification.

Open-weight

The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.

How tiers are computed

Public materials checklist

Items for a model under USASI rubric v0.2. Unknown means unassessed or insufficient evidence.

Public materials checklist for Granite 4.2 3B
Item	Status	Notes and evidence
WeightsCan the general public download the model parameters for this release?	Public	Safetensors weights in bfloat16 are published in the Hugging Face repository.1
Inference codeIs code for running the model published?	Public	The model card documents running the model with Hugging Face Transformers, vLLM, and SGLang, and the repository includes a chat template and a reasoning-parser plugin.1
Training codeIs the code used to train the model published?	Unknown	The model card names NVIDIA's NeMo RL and NeMo Gym for reinforcement learning. This review did not find IBM's own training code or configurations for this release.
Training-data informationDoes the information cover provenance, scope, acquisition, selection, labeling, processing, and where data or alternatives can be obtained? Access alone does not establish completeness.	Unknown	Not assessed.
Training-data accessCan the training data be obtained? This is independent of information completeness and reuse rights; original unshareable data need not be downloadable.	Unknown	Not assessed.
Complete training pipelineIs the complete base-training and preprocessing pipeline published, including configuration? Fine-tuning code or an inference SDK alone is insufficient.	Unknown	Not assessed.
Legacy data assessment (v0.1)Historical assessment combining download access and disclosure. Preserved for traceability; excluded from the v0.2 tier calculation. See the new separate assessments above.	Partial	IBM's disclosure file lists pre-training and post-training datasets by category, with URLs for most of them. Several entries are internal, acquired, or synthetic datasets that are not publicly obtainable.5
Training recipeAre the training configuration and procedure documented in enough detail to follow?	Partial	The technical blog documents the post-training procedure, including the SFT data mixture and configuration (learning rate, batch size, warm-up, epochs) and per-stage GRPO reinforcement-learning and RLHF settings. Pre-training of the Granite 4.1 base model is referred to a separate Granite 4.1 blog that this review did not assess, and no training code or configuration files were found.1 6
Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only.	Partial	The model card publishes a results table and states that evaluations use a framework based on NVIDIA's NeMo Evaluator SDK. This review did not find IBM's evaluation configurations or prompts.1

What it is useful for

The model card lists reasoning, code generation, tool calling, agentic workflows, and multilingual dialog in 12 tested languages.1

Run and use notes

Documented facts only. No hardware or performance claims are made without a cited source and stated assumptions.

The model card lists a 512K context window among key capabilities and describes native 128K support with long-context extension to 512K, but its architecture table gives a sequence length of 131,072, its vLLM and SGLang examples set a 131,072-token context limit, and the repository's config.json sets max_position_embeddings to 131,072 with no RoPE scaling. The card does not describe how to enable the 512K extension.1 2
The model card documents serving with vLLM (v0.20 or later for the bundled granite_thinking_parser) and with SGLang (v0.5.18 or later), and loading with Hugging Face Transformers. It recommends temperature 1.0 and top_p 0.95 for all modes.1
IBM also publishes quantized variants of this model in the ibm-granite Hugging Face organization, including GGUF, FP8, MXFP4, NVFP4, and MLX builds.7

Organization context

Provenance and derivatives

Post-trained by IBM from its own Granite 4.1 3B Base model, using supervised fine-tuning, reinforcement learning, and RLHF.1 6

Derived from: Granite 4.1 3B Base (external site: huggingface.co) — IBM base model; pre-trained from scratch per the Granite 4.2 technical blog.

Other releases in the IBM Granite family

Granite 4.2 30BModel-disclosure tier (USASI rubric v0.2): Open-weight
Granite 4.2 8BModel-disclosure tier (USASI rubric v0.2): Open-weight
Granite Switch 4.1 8BModel-disclosure tier (USASI rubric v0.2): Open-weight

IBM Granite family overview

In the news

Dated, sourced updates in this catalog's news that mention this entry.

Aug 25, 2026 · Release
IBM releases Granite 4.2 language models in 3B, 8B, and 30B sizes under Apache 2.0
IBM released Granite 4.2 on August 25, 2026: dense, decoder-only language models with built-in reasoning in 3B, 8B, and 30B sizes, post-trained from IBM's own Granite 4.1 base models. The weights are published on Hugging Face under the Apache 2.0 license. The catalog has a release record for each size under the Granite family.

What this catalog does not know

Unknown means the sources reviewed for this record do not document it. It is not evidence that something does not exist.

Training code: unknown.
Training-data information: unknown.
Training-data access: unknown.
Complete training pipeline: unknown.

Have a primary source? How to report a correction.

U.S. eligibility

Project eligibility rests on documented governing or maintaining entities, not on contributors.

Eligible · basis: U.S. headquarters

Developed by IBM's Granite Team per the model card. IBM's principal executive offices are in Armonk, New York, per its fiscal 2025 Form 10-K. It is post-trained from IBM's own Granite 4.1 3B Base, not from another developer's weights.1 8

Assessed Sep 29, 2026

Sources

1.
ibm-granite/granite-4.2-3b model card (external site: huggingface.co)
IBM (Hugging Face) · Model card · published Aug 25, 2026 · accessed Oct 2, 2026
2.
ibm-granite/granite-4.2-3b config.json (external site: huggingface.co)
IBM (Hugging Face) · Repository · accessed Oct 2, 2026 · evidence reviewed Oct 2, 2026
3.
granite-4.2-language-models LICENSE (external site: github.com)
IBM · License · accessed Sep 29, 2026
4.
Granite 4.2 disclosures README (external site: github.com)
IBM · Documentation · accessed Sep 29, 2026
5.
Granite 4.2 3B disclosure (JSON) (external site: github.com)
IBM · Documentation · published Sep 3, 2026 · accessed Sep 29, 2026
6.
Granite 4.2 LLMs: How They're Built (external site: huggingface.co)
IBM Granite (Hugging Face blog) · Release notes · published Aug 25, 2026 · accessed Sep 29, 2026
7.
Hugging Face model listing for ibm-granite (search granite-4.2) (external site: huggingface.co)
Hugging Face · Other · accessed Sep 29, 2026
8.
IBM Form 10-K for the fiscal year ended December 31, 2025 (external site: sec.gov)
U.S. Securities and Exchange Commission (filed by IBM) · Filing · published Feb 24, 2026 · accessed Sep 29, 2026

Email a correction Report a correction Edit this entry

This listing is not an endorsement, a safety assessment, or a federal approval.

Cite this page

APA style

USASI. (2026, October 2). Granite 4.2 3B. United States of America Superintelligence. https://unitedstatesofamericasuperintelligence.com/open/granite-4-2-3b/

BibTeX

@misc{usasi_granite_4_2_3b,
  author = {{USASI}},
  title = {Granite 4.2 3B},
  year = {2026},
  month = oct,
  howpublished = {\url{https://unitedstatesofamericasuperintelligence.com/open/granite-4-2-3b/}},
  note = {United States of America Superintelligence. Last updated 2026-10-02}
}

The date is when this page was last updated. Add the date you read it if your style needs one.

For maintainers: a “Listed in USASI” badge

Links to this page. It says the entry is listed, not endorsed or rated. About the badge

Markdown (for a README)

[![Listed in the USASI catalog](https://unitedstatesofamericasuperintelligence.com/badges/listed.svg)](https://unitedstatesofamericasuperintelligence.com/open/granite-4-2-3b/)

HTML

<a href="https://unitedstatesofamericasuperintelligence.com/open/granite-4-2-3b/"><img src="https://unitedstatesofamericasuperintelligence.com/badges/listed.svg" alt="Listed in the USASI catalog" height="20"></a>