granite-4.2-3B

About

Granite-4.2-3B is IBM Granite's compact reasoning model - a 3B decoder-only dense Transformer that delivers strong performance on reasoning-intensive tasks via built-in |-|...|-| chain-of-thought (the card's literal thinking delimiters). It is part of the Granite 4.2 generation, which introduces native reasoning (thinking) capabilities with flexible modes - full thinking (default), non-thinking, and low-effort - letting users balance depth versus latency per query.

Architecturally it uses Grouped-Query Attention (GQA) with 40 attention heads / 8 KV heads (head size 64), RoPE with theta 10,000,000, SwiGLU MLP (hidden 8192), RMSNorm, untied embeddings, 40 layers, embedding size 2560, and bfloat16 precision. It natively supports a 128K context window with a long-context extension to 512K, and was tested in English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese.

It was post-trained from Granite-4.1-3B-Base through a three-stage pipeline: SFT on instruction-following, chain-of-thought and reasoning data (public permissive datasets, internal synthetic reasoning/tool-calling data, agentic traces, human-authored data); then multi-phase, multi-environment reinforcement learning with Group Relative Policy Optimization (GRPO) across math, code, science, instruction following, tool use, general chat and structured output - most environments provide verifiable rewards (open-ended prompts scored by a generative reward model) - running asynchronous RL (generation and policy updates on separate GPU pools, weights refreshed in flight), followed by preference-alignment RLHF. RL ran on NeMo RL / NeMo Gym on an NVIDIA GB200 NVL72 cluster (CoreWeave).

Best for reasoning, code generation, tool calling, agentic workflows and multilingual dialog. Released August 25, 2026 under Apache 2.0.

Training Data Post-trained from Granite-4.1-3B-Base (pre-trained on large-scale English and multilingual corpora). Stage 2 SFT: public permissive datasets + internal synthetic reasoning/tool-calling/CoT data + agentic traces + human-authored data. Stage 3 RL: multi-phase multi-environment GRPO (math, code, science, IF, tool use, chat, structured output) with verifiable rewards, then preference-alignment RLHF; NeMo RL / NeMo Gym; GB200 NVL72 cluster on CoreWeave.