IBM has launched Granite 4.2, a new open-weight LLM family for self-hosting.
It comes in 3B, 8B, and 30B sizes.
The 8B and 30B versions get extra agentic training.
What happened
IBM has released its new Granite 4.2 family of open-weight large language models, available in 3B, 8B, and 30B parameter sizes. These models are designed to be downloaded and self-hosted.
Why it matters
The release is IBM's reasoning-focused update, featuring a native 128,000-token context window. The 8B and 30B variants also include agentic reinforcement-learning training for tools like terminal use and web search.
What to watch
The 3B model supports tools but lacks the specialized training of the larger variants, so businesses may need to evaluate which size fits their on-premise needs based on capability and resource trade-offs.
Ask the AI about this article →
IBM's Granite 4.2 release underscores a growing industry interest in local, self-hosted LLMs, appealing to businesses that want data privacy and control. The reasoning-focused approach, using chain-of-thought, is a key differentiator. The 8B and 30B variants' agentic training could enable more autonomous workflows, though the 3B model remains a lighter option. This suggests IBM is positioning Granite as a flexible solution for various deployment scenarios.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Anthropic's IPO is set to value the company at around $2 trillion, with OpenAI targeting a similarly large fig…
Caltech Professor Anima Anandkumar developed FourCastNet, an open-source AI weather model that rivals traditio…
Meta halted its planned AI-driven layoffs hours before the first round, after internal documents revealed a pl…
Goldman Sachs Global Institute co-head Jared Cohen argues that the world will be redesigned for machine intell…
Sentence Transformers v6.0 adds a fourth model type, MultiVectorEncoder, for ColBERT-style late-interaction re…
Perceptron, founded in November 2024 by former Meta FAIR scientists Armen Aghajanyan and Akshat Shrivastava, l…
