IBM Releases Granite 4.2 Open Reasoning Models with Agentic RL

Unlike earlier Granite releases, which were instruction-following assistants, Granite 4.2 is built around explicit reasoning. Every model can emit a chain of thought before answering and exposes a thinking / non-thinking switch plus a low-effort mode for easy questions. The models are decoder-only dense transformers, pre-trained from scratch on roughly 15 trillion tokens, then post-trained through a multi-stage reinforcement learning chain. For the 8B and 30B versions, the chain includes an agentic RL block where the model learns to edit code, drive a terminal, and run web searches in real sandboxed environments. The architecture supports a 128K sequence length, with pre-training extending to 512K tokens.

The models target different users. The 3B fits solo developers and startups running on a laptop through Ollama or LM Studio, with released GGUF quantizations down to Q4_K_M. The 8B suits mid-market teams on a single modern GPU. The 30B targets enterprises with A100/H100-class capacity, or FP8/NVFP4 serving on vLLM. Regulated organizations get an additional benefit of on-prem weights. Applications span software engineering agents, terminal and DevOps automation, deep-research and search agents, long-document RAG, structured tool calling, and high-volume transcription.

In reported benchmarks, the 30B model scored 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1. On AIME 2025, scores were 78.33 (3B), 86.67 (8B), and 89.17 (30B). The models also achieved 64.14 (8B) and 66.41 (30B) on GPQA, and 74.04 (8B) and 77.60 (30B) on MMLU-Pro. The agentic RL block runs only on the 8B and 30B, which explains most of the capability gap across sizes.

The two Granite Speech 5.0 Turbo CTC models, each with 470 million parameters, drop the LLM backbone entirely and use connectionist temporal classification to map audio to text. IBM reports an RTFx throughput near 12,600 on a single H200, against roughly 6,000 for current speed leaders on the Open ASR leaderboard. A WebGPU demo is live.

Training details include supervised fine-tuning on about 7.2 million samples, roughly 100 billion tokens, with a mixture of 31.6% agentic and 68.4% non-agentic data. Post-training is a multi-stage asynchronous GRPO run that warm-starts from the previous checkpoint, including RLVR, skill boosters, SWE, Terminal, Search, and RLHF. The training ran on NeMo-RL and NeMo-Gym over an NVIDIA GB200 NVL72 cluster hosted by CoreWeave. Also, 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline and a speculative decoding layer support the release.