IBM Granite 4.2 Ships Agentic RL Under Apache 2.0

IBM released Granite 4.2 on August 25, publishing the weights for three models — 3B, 8B and 30B — under an Apache 2.0 licence that permits commercial use without conditions. The headline change is not size. Every model in the family now carries a switchable thinking mode, and the two larger ones were put through reinforcement learning inside live software-engineering, terminal and web-search environments rather than trained purely on static text.

That distinction is the whole release. IBM calls the second stage agentic RL: instead of learning tool use from transcripts of tool use, the 8B and 30B models were dropped into sandboxes where they had to edit code, drive a terminal and run searches, and were rewarded on whether the task actually completed. The 3B skipped that stage, which IBM is candid about — it explains why the size gap on agentic benchmarks is wider than the parameter counts alone would suggest. Supervised fine-tuning drew on roughly 7.2 million samples before a sequence of GRPO phases covering reasoning, skill boosters and specialised tasks.

The scores are respectable rather than record-breaking, which is the point of a model you can run yourself. The 30B posted 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1, alongside 89.17 on AIME25 and 77.60 on MMLU-Pro. All three models were trained from scratch on about 15 trillion tokens with context windows reaching 512,000 tokens, tested across twelve languages including German, Japanese, Arabic, Korean and Chinese. Underneath, per IBM's model cards, the architecture is unshowy decoder-only plumbing — grouped-query attention, RoPE embeddings, SwiGLU MLPs in bfloat16, with 64 layers in the largest variant.

Deployment is where the sizes earn their keep. The 3B is meant to run on a laptop, the 8B on a single GPU, the 30B on enterprise infrastructure — a ladder aimed squarely at companies that want an agent touching internal systems without the traffic leaving the building. The weights are on Hugging Face, Ollama and GitHub, with OpenAI-format tool calling and support for vLLM and SGLang, so the serving stack most teams already run needs no adaptation.

IBM also shipped Granite Speech 5.0 Turbo, a 470-million-parameter CTC transcription model it claims runs roughly twice as fast as the previous benchmark leaders, at around 12,600 RTFx — fast enough, on IBM's numbers, to get through three hours of audio in a second. It is the least discussed part of the announcement and possibly the most immediately useful one, given how much enterprise transcription is still billed per minute.

The context is a crowded year for open weights. Moonshot put a 2.8-trillion-parameter model in the public domain in July, Mira Murati's Thinking Machines released the 975B Inkling a week earlier, and MiniMax M3 arrived with a million-token window. Granite 4.2 is not competing on any of those axes. It is competing on the boring question of what a regulated enterprise can legally download, audit and run on hardware it owns — and on that question, a 30B model with real agentic training and an Apache licence is a more interesting object than another frontier-scale checkpoint most companies will never self-host.