IBM released Granite 4.2 on August 25, a family of open-weight models that ships with a reasoning mode built in and, in the two larger sizes, training that taught the models to use tools inside real software environments. The release comes under an Apache 2.0 license, which means companies can download the weights, fine-tune them, and run them on their own hardware without paying IBM a cent.

Three sizes arrive together: a 3-billion-parameter model small enough for edge devices, an 8-billion mid-tier, and a 30-billion flagship that can stretch its context window to 512,000 tokens. Every model can switch between a thinking mode, where it works through a problem step by step, and a faster non-thinking mode for simpler tasks. That toggle matters in practice, because reasoning burns compute and most requests do not need it.

Trained to act, not just answer

The more interesting part is how IBM trained the 8B and 30B versions. Rather than only learning from text, they went through what IBM calls agentic reinforcement learning. The models practiced writing and running code, working in a terminal, and searching the web inside sandboxed environments, and they were rewarded for actually finishing the task. According to MarkTechPost, the pipeline starts from scratch on roughly 15 trillion tokens before that agentic phase begins.

The pitch is aimed squarely at businesses that want capable models running inside their own walls. A bank or a hospital that cannot send data to an outside API can put a Granite model on local servers, keep the information in house, and still get the tool use and reasoning that until recently meant reaching for a frontier lab's cloud.

The open field is getting crowded

Granite 4.2 lands in a year when open weights have stopped being an afterthought. Google's Gemma models just crossed a billion downloads, Z.ai has pushed GLM-5.3 deeper into coding, and Nvidia built a trillion-parameter open model of its own. IBM's angle is not raw size. It is the mix of small footprints, a permissive license, and agent skills tuned for the unglamorous work companies actually run.

Whether Granite can hold its own against those rivals on hard benchmarks is the open question. What is clear is that a company can pick it up today, read the license, and deploy it without asking anyone's permission. For a lot of enterprise buyers, that freedom is the feature.

Sources

  1. i. the-decoder.com
  2. ii. www.marktechpost.com
  3. iii. huggingface.co
  4. iv. www.unite.ai

Commentarii · 0

Add · a · Comment