IBM released Granite 4.2, its first family of dense, decoder-only reasoning large language models, across three parameter sizes of 3B, 8B, and 30B under the Apache 2.0 license, Hugging Face reported. Each model features a thinking and non-thinking switch, a low-effort mode for shorter reasoning budgets on simple queries, and native tool calling formatted for OpenAI-compatible endpoints.
The Granite team pre-trained the models from scratch on approximately 15 trillion tokens across a five-phase curriculum that extends the context window to 512,000 tokens. The underlying architecture uses Grouped Query Attention with 40 attention heads and eight key-value heads, Rotary Position Embeddings with a theta value of 10,000,000, SwiGLU activations, RMSNorm, and separate input and output embeddings executed in bfloat16 precision.
Supervised fine-tuning used approximately 7.2 million samples containing roughly 100 billion tokens, of which 65 billion were trainable. That dataset combined 31.6 percent agentic data—covering software engineering, terminal execution, and tool use—with 68.4 percent non-agentic conversational, coding, and mathematical data. The team filtered data using GPT-OSS-120B and Gemma 4 as judges alongside SHA-256 hash deduplication. For the 30B model, developers added a second supervised phase focused specifically on agentic coding, retaining about 16 percent replay data.
Reinforcement learning pipeline
Post-training uses asynchronous Group Relative Policy Optimization across NeMo-RL and NeMo-Gym to decouple rollout generation from trainer optimizer steps. Each stage warm-starts from the preceding checkpoint, beginning with verifiable-reward reinforcement learning across mathematics, Lean formal proofs, competitive programming, and structured outputs. Targeted booster stages follow to improve multi-turn instruction following and coding.
Training paths diverge after the foundational stages. The 3B model routes directly to alignment, while the 8B and 30B models complete an agentic reinforcement learning block in live environments. This block trains models on software engineering tasks using the OpenHands harness, live terminal operations using the Harbor and Terminus-2 harnesses across up to 64 environment turns, and multi-hop web browsing. Every model finishes with reinforcement learning from human feedback, applying a generative reward model, safety penalties, and a reasoning-length penalty to curb verbosity.
Hardware and deployment
CoreWeave hosted the training compute on an Nvidia GB200 NVL72 cluster linked by a 400 Gb/s InfiniBand fabric. For inference and deployment, the models run in vLLM and SGLang. IBM also released quantized variants in FP8, NVFP4, MXFP4, and multiple GGUF formats using llama.cpp.
