IBM releases Granite 4.2 in 3B, 8B, and 30B sizes

IBM Granite 4.2 adds reasoning and tool use across three open model sizes. Here is what launched, what changed, and how to evaluate it.

Quick answer

IBM released Granite 4.2 as 3B, 8B, and 30B open models focused on reasoning, native tool calling, coding, and agent workflows. The weights are available under Apache 2.0, but Granite 4.2 availability in Chat AI has not been verified.

Download Chat AI Opens the official App Store or Google Play for your device.

What IBM released

IBM announced three Granite 4.2 language models with 3B, 8B, and 30B parameters. IBM describes them as dense, open models designed for reasoning and agent workflows, and says the models are available through repositories and runtimes including Hugging Face, GitHub, Ollama, and other supported platforms.

Sources: IBM Research, IBM Granite on Hugging Face

Reasoning and tool use are the main changes

Every Granite 4.2 size can switch between thinking, non-thinking, and low-effort thinking modes and supports native tool calling. IBM says the 8B and 30B versions also received agent-focused reinforcement learning for tasks such as software engineering, terminal work, and web search, while the 3B version follows a shorter post-training path.

Sources: IBM Granite on Hugging Face

The three sizes serve different deployment needs

The shared family gives teams a direct size trade-off rather than a single fixed model. IBM positions smaller models for higher-throughput work and the larger model for deeper reasoning and more complex coding workflows. Actual speed, memory use, and answer quality will still depend on the hardware, quantization, serving stack, prompt, and task.

Sources: IBM Research, IBM Granite on Hugging Face

Open weights make independent testing possible

IBM released the Granite 4.2 language models under the Apache 2.0 license. That makes it possible to inspect the model cards, run the weights in a compatible environment, and test them on a real workload. Treat IBM's published benchmark results as first-party evidence, then verify instruction following, tool calls, accuracy, latency, and operating cost in the environment you plan to use.

Sources: IBM Research, IBM Granite on Hugging Face, IBM Granite on Hugging Face

Granite 4.2 is not yet verified in Chat AI

Granite 4.2 does not appear in Chat AI's public model directory as of 25 August 2026, so this article does not claim that the model is available in the app. Check the current directory for verified model access, which can vary by plan, platform, region, and app version.

Sources: Chat AI

Frequently asked questions

What readers usually ask

What is IBM Granite 4.2?

Granite 4.2 is IBM's family of 3B, 8B, and 30B language models focused on reasoning, instruction following, tool calling, coding, and agent workflows.

Is Granite 4.2 open source?

IBM describes Granite 4.2 as an open model release and publishes the language-model weights under the Apache 2.0 license. Review the license and model card for the exact files you plan to use.

Which Granite 4.2 size should I use?

Choose with a controlled test. A smaller model may fit tighter hardware or higher-throughput work, while the 30B model is positioned for harder reasoning and coding tasks. Measure quality, latency, memory use, and cost in your own environment.

Is Granite 4.2 available in Chat AI?

Granite 4.2 availability in Chat AI was not verified at publication time. Check the current Chat AI model directory before assuming access.

Evidence

Sources

  1. Granite 4.2 release announcementIBM Research · Primary source
  2. Granite 4.2 technical walkthroughIBM Granite on Hugging Face · Primary source
  3. Granite 4.2 30B model cardIBM Granite on Hugging Face · Primary source
  4. Chat AI model directoryChat AI · Primary source