TL;DR — Key Takeaways
- IBM Granite 4.2 adds an optional reasoning mode across its 3B, 8B and 30B open-weight models.
- The 8B and 30B models received additional reinforcement learning focused on agentic tasks such as coding, terminal use and web search.
- IBM trained the models on 1 trillion tokens of synthetic code, while the 30B model received extra tuning for software engineering and agentic coding.
IBM announced the release of Granite 4.2, an update to its open-weight family of 3B, 8B and 30B language models.
The three models are available under the Apache 2.0 license through platforms including Hugging Face, GitHub, Ollama and IBM watsonx. Granite 4.2 adds the option for models to reason through a problem before producing an answer, while still letting users turn off that mode for simpler tasks.
Adding a dedicated reasoning mode is notable because IBM had recently touted the efficiency of handling many enterprise workloads without that extra computational overhead. When the company released Granite 4.1 in April, it said the models could deliver competitive instruction-following and tool-calling performance without the additional token use and latency of long chains of thought.
With Granite 4.2, IBM is layering more reasoning and agent-specific training onto the tool-use foundation established with 4.1. The 4.2 models underwent supervised fine-tuning followed by foundational reinforcement learning focused on areas including math, science, coding, reasoning and tool use. IBM said the foundational RL stage combined objective checks, such as whether code passed tests or an answer matched the correct result, with reward-model evaluation for more open-ended responses.
For the two larger models, IBM went a step further. The 8B and 30B models received an additional reinforcement learning stage in which they practiced tasks like editing and running code, using a terminal and searching the web in sandboxed environments. IBM said the extra training was aimed specifically at the kinds of multi-step tasks AI agents are now expected to perform. The 3B model did not receive that agentic RL stage, but it still supports tool calling and received the foundational RL applied across the family. IBM also placed additional training emphasis on coding, which has been an area of rapid adoption for agentic AI. The models were trained on 1 trillion tokens of synthetic code generated through IBM’s CodeAlchemy pipeline, while the 30B version received an additional fine-tuning stage weighted toward software engineering and agentic coding tasks. In IBM’s testing, the 30B model resolved 57% of problems on SWE-bench Verified, a benchmark that tests models against real-world software issues drawn from GitHub repositories.
Granite 4.2’s open release also allows for more control over where and how the models run. They can be deployed in cloud, on-premises, and edge environments alike. Because the weights are available under the Apache 2.0 license, organizations can download, fine-tune and self-host the models without exclusive reliance on a hosted API. Those more flexible deployment options can appeal to organizations handling sensitive data, navigating data residency requirements or looking to keep AI workloads within on-prem infrastructure they control.
The new release suggests IBM is trying to balance the demands of more complex agentic workloads with its earlier focus on inference efficiency. With Granite 4.2, IBM is following the industry’s shift toward reasoning-heavy agents but preserving its emphasis on openness and deployment flexibility.