IBM has released its Granite 4.2 language models in 3 billion, 8 billion and 30 billion parameter sizes. The models were trained from scratch on about 15 trillion tokens.

The most notable part of the release is the license: all models ship under Apache 2.0. That is a permissive license covering commercial use, which for an enterprise buyer means the right to run the model on their own infrastructure, modify it and embed it in a product.

Technical specifications

  • Context window — up to 512,000 tokens.
  • Thinking mode — the models can toggle between "thinking" and "non-thinking" modes to control compute per task. A "low-effort" mode saves resources on simple queries.
  • Agentic training — the 8B and 30B variants go through what IBM calls "agentic RL": 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 serving stacks.

The speech side

The new speech models shipping alongside carry just 470 million parameters. According to IBM they are twice as fast as the previous leaders on the open speech recognition leaderboard and can transcribe three hours of audio in one second.

The figure is striking but belongs in context: speech recognition is a far narrower task than language generation, and small models performing well there is expected. Even so, a 470-million-parameter model running at that speed marks a meaningful threshold for applications running on device.

Why these sizes

The three sizes are not arbitrary. Each maps to a different hardware class: the 3-billion-parameter model runs on a laptop, the 8-billion on a single server GPU, and the 30-billion on a modest server cluster.

That is a different choice from the path frontier models follow. Frontier labs ship one enormous model and sell access through an interface; IBM offers a range an organization can pick from based on the hardware it already has. The two answer different questions: one is "what is the best," the other is "what is the best that runs here."

Why it matters

The release lands amid rising interest in locally-run language models, and IBM's positioning targets that directly.

For an enterprise buyer the arithmetic works like this. With a closed model you pay per token, your data goes to the provider, and the price is under the provider's control. With an Apache 2.0 model of 8 billion parameters running on your own servers, all three disappear. In exchange the model is weaker than frontier systems — but many enterprise tasks do not require frontier capability.

The sizes IBM chose support that reading. 3B, 8B and 30B are scales that run on a laptop, a single server GPU and a modest server cluster respectively. The company is not competing at the frontier; it is selling a layer meant to run inside an organization's own walls.