IBM has released the Granite 4.2 language model family. It comprises three sizes with around 3, 8 and 30 billion parameters and is fully available under the Apache 2.0 licence – a free licence that expressly permits commercial use, redistribution and modification. According to IBM, around 15 trillion tokens were used for training; one token corresponds to a word or part of a word.

Notable is the context window of up to 512,000 tokens. The context window describes how much text a model can take into account simultaneously in a single pass. At this scale, extensive document collections, longer log excerpts or larger parts of a codebase can be processed in one go without having to split them into small chunks beforehand.

Agentic training as the real innovation

The two larger models were additionally trained with so-called agentic RL – reinforcement learning in which the model not only predicts text but is rewarded for successfully completed action chains. Specifically, this involves independently calling tools and executing code.

The difference from a classically trained model is practically relevant. A pure text model can describe which API call would make sense. A model trained on tool use, by contrast, decides during the process when a call is necessary, shapes the parameters, evaluates the response and derives the next step from it. This is exactly where many agent prototypes fail: not on language understanding, but on the reliability of the intermediate steps.

Why open weights matter for mid-sized companies

Open weight means that the trained model parameters can be downloaded and operated on your own infrastructure. Together with the Apache 2.0 licence, this results in several effects that regularly play a role in project decisions:

  • Data sovereignty: Content does not necessarily leave your own environment. This simplifies arguments vis-à-vis data protection officers and clients in regulated industries.
  • Predictable costs: Instead of billing per token, hardware and operating costs apply. With consistent load this is easier to calculate; with strongly fluctuating usage it is not automatically cheaper.
  • No dependency on a model version: A locally operated model is not switched off or have its behaviour changed without warning. For productive automations, this is an underestimated stability factor.
  • Adaptability: Fine-tuning to your own specialist language, product catalogues or internal documentation structures remains possible.

The size gradation is an architecture decision

Releasing three model sizes is no coincidence. In practice, most tasks in a software project do not need the largest available model. Classifying incoming requests, extracting structured fields from forms or summarising short texts run quickly and resource-efficiently with smaller models – a 3B model can be operated on comparatively moderate hardware.

More demanding steps, in which several tools must be coordinated in sequence, are the domain of the larger variants with agentic training. Anyone who plans from the outset for different tasks to go to differently sized models keeps response times and operating costs under control without sacrificing quality at the critical steps.

What this means for concrete projects

For web and software projects, several obvious fields of application emerge. In the environment of a content management system such as TYPO3, editorial tasks can be supported: suggestions for metadata, checking internal links, preparing content for additional output channels. A large context window helps where it is not a single text but the structure of an entire page tree or an extensive migration inventory that must be considered.

In the development process itself, the ability to execute code is interesting. An agent that not only writes a script but runs it on a test basis, reads the error pattern and corrects it shortens feedback loops in routine work such as data cleansing, format conversion or recurring evaluations.

With process automation, it is usually not about impressive individual responses but about reliability across many runs. That is precisely why training on action chains is a more sensible selection criterion than pure benchmark values.

Stay level-headed in assessment

Open weights do not solve all questions. Self-operation shifts effort: hardware procurement, model deployment, monitoring, updates and security then lie within your own area of responsibility. Anyone without experience in operating GPU-backed services today should budget this effort realistically.

Equally important: a model that may call tools and execute code needs clear boundaries. Sandboxing of the execution environment, restrictive permissions on connected systems, logging of all calls and a defined abort mechanism are not optional refinements but prerequisites for productive use. Agentic capabilities increase both the benefit and the potential for damage at the same time.

A delimited first use case with a measurable result makes sense – for example a recurring manual work step whose duration and error rate are known today.

Only when reliable figures are available there is it worth expanding.

Classification

Freely licensed models with agentic training shift the question of whether AI automation is feasible without the major cloud providers from a fundamental one to an operational one. For projects where data sovereignty, calculable costs or long-term version stability matter, this creates a genuine alternative to API integration. What remains decisive is the architecture around it: clearly defined tasks, appropriately sized models and hard technical limits for everything an agent may execute autonomously.

Sources