Granite 4.2

What Granite 4.2 is

Granite 4.2 is a generation of open-weight language models from IBM's Granite team, listed with a release date of 25 August 2026^[1]. It comes in three dense sizes: 3B, 8B and 30B^[1]. "Dense" means the whole parameter count is used to process every token, unlike mixture-of-experts designs that send each token through a small subset of the network; the size numbers therefore map fairly directly onto the cost of running the model. A team that wants a model on a single workstation can take the small one, while a team serving many users can take the large one, with the same family and licence across the range.

Reasoning inside the 8B model

The most detailed public information concerns the 8B member. It is built on Granite-4.1-8B-Base and adds native chain-of-thought reasoning inside think tags^[1]. Chain of thought means the model writes out intermediate steps before committing to a final answer. Putting those steps inside dedicated tags lets software separate the working from the answer: an interface can hide the scratchpad from the reader, and a logging system can keep it for checking.

The model offers three thinking modes: full thinking, non-thinking and low-effort^[1]. The point of the switch is that reasoning costs tokens and time. A hard problem may justify long internal working; a routine request such as reformatting a document usually does not, and answering it directly is cheaper. Low-effort sits between the two, giving the model room for a short amount of working without the full expense.

Architecture of the 8B model

The 8B model has 40 layers with an embedding size of 4,096^[1]. Its attention uses grouped-query attention with 32 attention heads and 8 key-value heads^[1]. In grouped-query attention, several query heads share a single key-value head; this shrinks the stored key-value cache that grows with every token of context, which is what makes long conversations and long documents practical to serve. The position embedding is RoPE with a theta of 10,000,000^[1]; rotary position embedding encodes where a token sits by rotating its vector by an angle that depends on position, and the large theta value spreads those rotations out so that positions far apart stay distinguishable. The feed-forward layer uses SwiGLU with a hidden size of 12,800^[1], a gated variant of the usual feed-forward block. Input and output embeddings are not tied, which means the matrix used to read tokens in and the one used to produce them out are separate parameters, and the model runs in bfloat16^[1], a 16-bit floating-point format that keeps a wide exponent range for stability during inference.

Context is natively 128K tokens and extensible to 512K^[1]. A token is a word fragment; 128K tokens is on the order of a long book, which is what allows whole code repositories or document sets to be passed in at once. Extending the window to 512K pushes into the range of many books.

Languages and licence

Twelve languages are tested: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese^[1]. The list mixes large European and Asian languages with Czech, a smaller one, and Arabic written in a different script, which matters for organisations operating across several markets.

The licence is Apache-2.0, allowing unrestricted commercial and academic use^[1]. For a company, this removes the negotiation and usage caps that often come with open-weight releases, and it allows the weights to be modified and redistributed.

Release timeline

Weights were uploaded to Hugging Face on 7 August 2026, ahead of the stated release date^[2]. That gap matters to practitioners: people can download and test a model before its official announcement, so early reports tend to come from users rather than from the vendor. GGUF and NVFP4 builds followed in the same month^[2]. GGUF is a file format used by local inference software, which makes a model runnable on consumer hardware; NVFP4 is a 4-bit floating-point format, and a build in it shrinks memory use substantially at some cost in precision. Both builds point at the same audience: teams that want to run the model on their own machines rather than through an API.

Where it sits

Granite 4.2 continues an established pattern in open-weight releases: a family spanning small to mid-size dense models, a permissive licence, and quantised builds published alongside the full-precision weights so the model can leave a data centre. The addition of switchable reasoning modes is the notable part. It acknowledges that a single model now has to serve two quite different jobs — fast routine text handling and slower multi-step problem solving — and lets the caller decide which one a given request needs.