IBM launches Granite 4.2 with open-source chain-of-thought models

IBM released the Granite 4.2 family, three language models of 3 billion, 8 billion and 30 billion parameters built around explicit reasoning (Chain-of-Thought) rather than instruction following only. Each model can emit a reasoning chain before the answer, and all expose a switch between reasoning mode and normal mode plus a “low-effort” mode that allocates a short reasoning budget for simple queries. The models are dense decoder-only transformers, not MoE or hybrid architectures, and were trained from scratch on about 15 trillion tokens.

The three models are distributed under the Apache 2.0 license, allowing download, fine-tuning and commercial production use without licensing restrictions. The 3 B version targets individual developers and startups running on a laptop via Ollama or LM Studio, especially with GGUF quantizations down to Q4_K_M. The 8 B version is aimed at mid-market teams on a single modern GPU, and the 30 B version is intended for organizations with A100/H100 capacity or FP8/NVFP4 serving on vLLM. Regulated entities also gain the option to run the weights on-prem.

The core of the effort lies in the post-training pipeline. Supervised fine-tuning (SFT) used roughly 7.2 million samples, about 100 billion tokens with roughly 65 billion trainable parameters. The mix comprised 31.6 % agentic samples and 68.4 % non-agentic, with software engineering accounting for 69 % of the agentic portion. Trajectories were generated in environments that include OpenHands, SWE-agent, Terminus-2, MiniSWE, Codex and Goose. Quality control was performed by GPT-OSS-120B and Gemma 4 as judges, supplemented by SHA-256 de-duplication on tool fields and messages.

The post-training phase is a multi-stage, environment-aware RL chain rather than a single pass. Each stage runs a separate asynchronous GRPO starting from the previous checkpoint, using a leave-one-out baseline instead of a value network and truncated importance sampling to bound off-policy deviation. The order is RLVR, then skill boosters, then SWE, terminal, search, and finally RLHF. The agentic RL block runs only on the 8 B and 30 B models; the 3 B receives only basic RL and alignment, a design choice that explains most of the capability gap across sizes. Training ran on NeMo-RL and NeMo-Gym over a NVIDIA GB200 NVL72 cluster hosted by CoreWeave.

Two supporting components complete the stack: a trillion tokens of synthetic code from IBM’s CodeAlchemy pipeline, and a speculative decoding layer for faster model inference. On the speech side, IBM released two Granite Speech 5.0 Turbo CTC models of 470 million parameters each; they forgo an LLM backbone and use a direct CTC architecture for high-throughput transcription. IBM published benchmark results for each size, though the metrics have not yet undergone independent external validation.