IBM has released its Granite 4.2 language models in 3, 8 and 30 billion parameter sizes. The models were trained from scratch on about 15 trillion tokens and, according to the company, support context windows up to 512,000 tokens.

What stands out

  • Controllable compute — the models toggle between "thinking" and "non-thinking" modes, letting compute per task be tuned. There is also a "low-effort" mode that saves resources on simple queries.
  • Agent training — the 8B and 30B variants go through what IBM calls "agentic RL" training, in which they learn to use tools, write and run code, and search the web in real sandbox environments.
  • Compatibility — all models support OpenAI-format tool calling and run on common inference servers.

The same announcement refreshed the speech models. The new Granite Speech 5.0 Turbo models have just 470 million parameters and are, IBM says, twice as fast as the previous leaders on the open speech recognition leaderboard. By the company's measure they can transcribe three hours of audio in one second.

The size choice is meaningful too. None of the three aims to compete with the giant frontier models; 30 billion parameters counts as mid-range at today's scale. IBM is aiming somewhere else: models that can run on your own hardware, be audited, and carry a clear license.

The license is the real story

All the models are distributed under Apache 2.0 and available on common model distribution platforms.

That detail matters more than the specifications. Most open-weight releases arrive under custom licenses that restrict commercial use or require permission. Apache 2.0 is permissive: you can take the model, modify it, embed it in your product and sell it without having to open your own source.

For an enterprise buyer that is where the difference lies. License ambiguity is the real reason many AI projects never get past the legal department.

One more detail: a permissive license also clears the way for the model to be fine-tuned and redistributed. That matters for languages with relatively small data volumes — a small team can take the model, adapt it with its own data and share the result.

Why it matters

Granite 4.2 lands on top of growing interest in language models that run locally. The small sizes are a deliberate choice here: models at 3 and 8 billion parameters can run on an ordinary laptop or a modest server.

What that means is not only that the token bill disappears. Data stays in-house, the provider cannot change the price, and the model does not get retired one day. In regulated sectors those three often outweigh a benchmark score.

Baking agent training into the model is also notable. Until now, getting a model to use tools was the job of the harness written around it. A model that learned tool use in real environments during training may render part of that harness unnecessary — which for small teams is a direct saving in setup cost.