What makes Granite 4.2 different from earlier Granite models?

Granite 4.2 is IBM's first reasoning-focused release. It adds explicit chain-of-thought reasoning and a thinking/non-thinking switch, allowing the model to deliberate before answering.

Are all Granite 4.2 models the same?

No. The three sizes (3B, 8B, and 30B) share the same architecture, but only the 8B and 30B models undergo agentic reinforcement learning, which teaches them to operate as agents with tools.

Large Language ModelsAI Coding AssistantsOpen-Source AIHugging Face BlogAug 25, 2026Aug 26, 2026, 01:01 JST

IBM Releases Granite 4.2 Reasoning LLMs

Send on LINE

3 Key Points

What happened
IBM released Granite 4.2, its first family of reasoning large language models (LLMs, AI systems that understand and generate text), in three sizes: 3B, 8B, and 30B. The models are pre-trained from scratch on roughly 15T tokens and extend the context window to 512K tokens.
Why it matters
The 8B and 30B models add agentic training, meaning they can operate as agents that use tools, edit and run code, and search the web. All models include a thinking/non-thinking switch and native tool calling, making them practical for business automation and complex tasks.
What to watch
The models are released under the Apache 2.0 license, and the larger models (8B and 30B) can act as agents inside real sandboxed environments, which could enable more autonomous AI workflows.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

Earlier Granite releases were strong instruction-following assistants, but Granite 4.2 adds explicit reasoning capabilities. The training pipeline is extensive: pre-training on 15T tokens, supervised fine-tuning with chain-of-thought data, and a multi-stage reinforcement learning process. The 8B and 30B models receive an additional agentic RL block that enables them to use tools in real sandboxed environments, a significant step beyond typical LLM releases. This means businesses can deploy models that not only reason but also act, taking actions like writing code or searching the web, all under a permissive license.

FAQ

What makes Granite 4.2 different from earlier Granite models?: Granite 4.2 is IBM's first reasoning-focused release. It adds explicit chain-of-thought reasoning and a thinking/non-thinking switch, allowing the model to deliberate before answering.
Are all Granite 4.2 models the same?: No. The three sizes (3B, 8B, and 30B) share the same architecture, but only the 8B and 30B models undergo agentic reinforcement learning, which teaches them to operate as agents with tools.

Hugging Face BlogRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.