IBM released Granite 4.2, a family of three dense, decoder only reasoning models at 3B, 8B, and 30B parameters, all licensed under Apache 2.0. IBM built each size on roughly 15 trillion tokens of pre-training and gave every model a thinking and non-thinking switch that lets a developer trade latency for reasoning depth on a per-query basis.
The release matters less for its benchmark scores than for where IBM spent its post-training budget. Open weight labs have spent the past two years converging on similar chat and instruction following quality. Granite 4.2 is IBM’s attempt to differentiate on what the model does after it answers: call a tool correctly, edit a repository, drive a terminal session, and recover when the first attempt fails.
That differentiation shows up in the training pipeline. IBM fine-tuned the base models on a corpus that is roughly 32 percent agentic data (software engineering, tool calling, terminal use, math, search), then ran a multi-stage reinforcement learning process built on GRPO (Group Relative Policy Optimization), the same family of algorithm DeepSeek popularized for verifiable-reward training. The 8B and 30B models go through an additional agentic block that the 3B model skips entirely: three sequential stages named SWE agent, Terminal agent, and Search agent, each rewarded only when the model actually solves a task inside a real, non-simulated environment rather than producing a plausible-looking answer.
That distinction between staged practice and live problem-solving is the part worth noting. In the SWE agent stage, the model works inside a real repository sealed in its own container, and IBM’s OpenHands harness scores it purely on whether the hidden test suite passes once it finishes editing. The Terminal agent stage puts the model in a live shell through a Harbor and Terminus-2 harness, judging it only on whether the command sequence actually finished the job; IBM let this stage run the longest rollouts in the whole pipeline, letting the model take dozens of turns to plan, execute, and recover from mistakes. The Search agent stage grades multi-hop web research through an AI judge model instead of a fixed answer key, because there is no single correct string to check research against.
IBM reports that the 30B model scores 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1, both improvements over the 8B model’s 47.67 and 20.56 on the same benchmarks. Those figures come from IBM’s own Granite blog post, not an independent evaluation, and the post does not include results from external leaderboards or third-party replications. The 3B model, which never enters the agentic RL block, is not scored on either benchmark at all, an implicit admission that agentic behavior does not emerge from scale alone at that size.
All three models support native tool calling through an OpenAI-compatible endpoint, so they plug into harnesses like OpenCode, Pi, and OpenHands without custom adapters, according to IBM. The company also shipped FP8, NVFP4, MXFP4, and GGUF quantized variants for local and edge deployment. IBM trained the family on CoreWeave-hosted infrastructure built from NVIDIA’s GB200 NVL72 systems, the same class of hardware frontier labs are running for their own next training cycles.
For teams evaluating open weight models for coding or research agents, the agentic RL curriculum is now a more relevant comparison point than aggregate MMLU or Arena scores. Anyone building an agent on Llama, Qwen, or Mistral weights should ask whether that model’s post-training included reward signals from real tool use, or only from chat preference data, before assuming it will hold up inside a multi-turn harness.
Reported by IBM’s Granite team in a technical blog post published on Hugging Face on August 25, 2026.
