Granite 4.2

IBM's first dense reasoning Granite family: 3B, 8B and 30B, switchable thinking, tool calling, Apache 2.0 for enterprise use.

In short. Granite 4.2 is IBM's set of three open AI models built for business use. They can think step by step before answering, or skip that for speed, and they can call tools. The sizes run from a 3-billion-parameter model for small machines to a 30-billion-parameter model for a single big GPU. All are Apache 2.0.

Summary

IBM published Granite 4.2 on 25 August 2026. It is the company's first family of dense, decoder-only reasoning models, in three sizes: 3B, 8B and 30B. They were post-trained from the Granite 4.1 base models, which IBM says were pre-trained from scratch on about 15 trillion tokens, then fine-tuned on reasoning and agent data and refined with a multi-stage reinforcement-learning pipeline.

Every model has a thinking and a non-thinking mode, plus a "low-effort" mode that spends a short reasoning budget on easy questions, and native tool calling in the OpenAI function-calling format. The 8B and 30B models also go through an agentic reinforcement-learning stage in which they learn to use tools, edit and run code, drive a terminal and search the web inside sandboxed environments. The context length is 128K natively, extended to 512K for long-context use. IBM lists English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese as tested languages.

The BF16 checkpoints are about 59 GB for 30B, 18 GB for 8B and 7 GB for 3B. IBM also publishes FP8, MXFP4 and GGUF builds; the 30B MXFP4 version is about 17 GB. The license is Apache 2.0, and IBM says all Granite models carry it.

Good for

  • Enterprise assistants, retrieval-augmented generation and tool-calling agents where a clear Apache 2.0 license matters
  • Running a capable reasoning model on one GPU (30B MXFP4 is about 17 GB), a workstation (8B) or a small device (3B)
  • Switching reasoning on and off per request to balance quality and latency
  • Multilingual dialog in the 12 tested languages

Watch out for

  • Text only; there is no vision or audio input in these models
  • The 3B model skips the agentic training stage that the 8B and 30B receive
  • Languages beyond the 12 tested ones may work but have not been fully tested
  • Quality and benchmark claims are IBM's own; compare on your tasks against Qwen, Gemma and others

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the 3B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
3B BF16 3.7B 8.51 GB 4.25 GB 2.55 GB Laptop or phone
8B BF16 8.8B 20 GB 10 GB 6.07 GB Laptop or phone
30B BF16 29.3B 67 GB 34 GB 20 GB Gaming GPU
30B MXFP4 MXFP4 29.3B 67 GB 34 GB 19 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Granite 4.2?
Granite 4.2 is IBM's set of three open AI models built for business use. They can think step by step before answering, or skip that for speed, and they can call tools. The sizes run from a 3-billion-parameter model for small machines to a 30-billion-parameter model for a single big GPU. All are Apache 2.0.
Can I use Granite 4.2 commercially?
Yes. Granite 4.2 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Granite 4.2 open source?
Granite 4.2 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Granite 4.2 on my own computer?
Yes. The smallest variant, 3B, needs about 2.55 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Granite 4.2 released?
Granite 4.2 was released by IBM on August 25, 2026.
How long a context window does Granite 4.2 have?
Granite 4.2 supports a context window of up to 128K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 8, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
MiniCPM5-2B OpenBMB 1.08B – 2.52B 128K Apache 2.0 (Permissive) 0.745 GB Laptop or phone
gpt-oss OpenAI 21B – 117B 5.1B active 128K Apache 2.0 (Permissive) 15 GB Laptop or phone
Olmo 3.1 Allen Institute for AI (Ai2) 7.3B – 32.2B 64K Apache 2.0 (Permissive) 5.04 GB Laptop or phone
TwIL webAI 3.08B – 8.79B 128K webAI Non-Commercial (Non-commercial) 2.11 GB Laptop or phone