IBM Releases Granite 4.2, Its First Reasoning-Focused LLM Family in 3B, 8B, and 30B Sizes

IBM Details the Build Process Behind Granite 4.2

IBM has released Granite 4.2, described by the company as its first family of dense, decoder-only reasoning LLMs, available in three sizes: 3B, 8B, and 30B parameters. All models are released under the Apache 2.0 license and are supported through OpenAI-compatible endpoints (via vLLM) and SGLang.

Architecture

All three sizes share the same design: a dense decoder-only transformer using Grouped Query Attention (40 attention heads, 8 KV heads), Rotary Position Embeddings with θ = 10,000,000, SwiGLU-activated MLPs, RMSNorm (ε = 1e-5), untied input/output embeddings, and bfloat16 precision. The 3B model uses 40 layers with a 2560 embedding size; the 8B model uses 40 layers with a 4096 embedding size; the 30B model scales to 64 layers with a 4096 embedding size and a 32,768 MLP hidden size. The published architecture table lists a 131,072-token sequence length, though IBM states that a dedicated long-context training phase extends the effective context window to 512K tokens.

Pre-Training

Each model is pre-trained from scratch on approximately 15 trillion tokens using a five-phase strategy. Phases 1–2 cover foundational pre-training, phases 3–4 apply progressively higher-quality data annealing, and phase 5 handles long-context extension. According to IBM, the data blend and phase schedule closely follow the Granite 4.1 recipe.

Supervised Fine-Tuning

Post-pretraining, models go through SFT on a mixture of 7.2 million samples (roughly 100 billion tokens, ~65 billion trainable), split between agentic data (31.6%) and non-agentic data (68.4%). The agentic corpus is dominated by software engineering trajectories (69%), followed by tool calling (12.1%), terminal use (8.0%), math (3.5%), search (0.8%), and action tasks (0.2%), generated using scaffolds including OpenHands, SWE-agent, Terminus-2, Gemini CLI, Codex, and Goose. The non-agentic mixture spans instruction following (18.8%), coding (18.8%), math (14.6%), multilingual (7.0%), science (5.4%), reasoning (3.0%), and safety (0.8%).

IBM applied multi-stage quality filtering, using GPT-OSS-120B and Gemma 4 as LLM judges to remove low-scoring or hallucinated samples and invalid tool calls, followed by SHA-256-based deduplication across the corpus. Training used 32–128 nodes (scaled by model size) with 4x Grace/GB200 GPUs per node, a packed sequence length of 131,072 tokens, a global batch size of 128, a constant learning rate of 1.0e-5 after warm-up (3.0e-6 in Phase 2), and roughly two epochs.

The 30B model receives an additional SFT phase focused on agentic coding, upsampling SWE and agentic data while retaining 16% replay data from the original mixture, trained for one more epoch at a 3.0e-6 learning rate.

Reinforcement Learning and Capabilities

After SFT, all models undergo a multi-stage RL pipeline. The 8B and 30B models additionally receive agentic RL training inside sandboxed environments, teaching them to call tools, edit and run code, operate a terminal, and search the web. Every model in the family includes a thinking/non-thinking toggle, a low-effort reasoning mode for simpler queries, and native tool calling in OpenAI function-calling format. IBM has not disclosed benchmark scores or pricing for Granite 4.2 in this post.

What This Means

Granite 4.2 marks IBM's push into reasoning-native, agent-capable open models rather than pure instruction-followers. The size-tiered agentic RL split — reserving tool-use training for the 8B and 30B models while keeping the 3B model reasoning-only — signals a deliberate cost/capability tradeoff aimed at enterprise deployment where smaller models handle routine tasks and larger ones run in agentic pipelines. The Apache 2.0 license and OpenAI-compatible tool-calling format make the models straightforward to slot into existing agent harnesses, but without published benchmark results, independent verification of IBM's reasoning and agentic claims will have to wait for third-party evaluation.