IBM Granite 4.2 Release Notes

IBM has released Granite 4.2, a family of dense, decoder-only reasoning LLMs available in 3B, 8B, and 30B parameter sizes. These models are designed for explicit reasoning, supporting a chain-of-thought process that can be toggled between thinking, non-thinking, and low-effort modes, and are released under the Apache 2.0 license.

Model Architecture and Pre-training

Granite 4.2 models utilize a decoder-only dense transformer architecture. Key technical specifications include Grouped Query Attention (GQA) with 40 attention heads and 8 KV heads, Rotary Position Embedding (RoPE) with θ = 10,000,000, and SwiGLU activation in the MLP. The models are trained in bfloat16 precision.

Component 3B Dense 8B Dense 30B Dense
Embedding size 2560 4096 4096
Number of layers 40 40 64
Attention head size 64 128 128
Number of attention heads 40 32 32
Number of KV heads 8 8 8
MLP hidden size 8192 12800 32768
Sequence length 131,072 131,072 131,072

Pre-training was conducted from scratch on approximately 15 trillion tokens using a five-phase strategy. This process included foundational pre-training, mid-training with data annealing, and a final phase that extended the context window to 512K tokens.

Supervised Fine-Tuning (SFT)

SFT transforms the base models into instruction-following and reasoning assistants using a mixture of approximately 7.2 million samples (roughly 100B tokens). The data is split into:

  • Agentic Corpus (31.6%): Focused on software engineering (69%), tool calling (12.1%), terminal use (8.0%), math (3.5%), search (0.8%), and action (0.2%).
  • Non-Agentic Corpus (68.4%): Including instruction following (18.8%), coding (18.8%), math (14.6%), multilingual (7.0%), science (5.4%), reasoning (3.0%), and safety (0.8%).

Quality control involved normalization to OpenAI Chat format, LLM-based judging via GPT-OSS-120B and Gemma 4, and SHA-256 based deduplication. The 30B model underwent an additional second phase of SFT specifically for agentic coding, upsampling SWE and coding data for one additional epoch.

Multi-Stage Reinforcement Learning Pipeline

Granite 4.2 employs a multi-stage RL pipeline using asynchronous Group Relative Policy Optimization (GRPO). This method uses a leave-one-out baseline to remove the need for a separate value network. The pipeline is a sequence of warm-starts where each stage's policy becomes the base for the next.

Foundational RL

All model sizes undergo foundational RL, which includes:

  • RLVR (Verifiable-Reward RL): The broadest stage, covering math (including Lean formal proving), competitive coding, STEM MCQA, instruction following, and reasoning puzzles. RLVR runs for two rounds on 3B/8B and three rounds on 30B.
  • Skill Boosters: Targeted runs to sharpen instruction following (multi-turn chat, structured outputs) and competitive coding.

Agentic RL (8B and 30B only)

The 8B and 30B models undergo an additional agentic block to learn tool interaction in real environments:

  • SWE Agent: Uses the OpenHands harness to edit code and run tests in sandboxed repositories, rewarded by whether hidden tests pass.
  • Terminal Agent: Operates in a live shell via the Terminus-2 harness, with rollouts spanning up to 64 environment turns.
  • Search Agent: Performs multi-hop research using live web-search tools, rewarded by an LLM judge.

Alignment (RLHF)

Every model concludes with RLHF for human preference and safety, utilizing a generative reward model (GenRM) and a safety reward. This stage also applies a reasoning-length penalty to reduce verbosity.

Infrastructure and Deployment

Training was performed on an NVIDIA GB200 NVL72 cluster. The software stack utilizes NeMo-RL for the GRPO loop and NeMo-Gym to orchestrate rollouts and host sandboxed resources.

Quantization and Serving

Four quantized variants are available for vLLM inference:

  • FP8: Dynamic per-channel weights and per-token activations.
  • FP4: NVFP4 and MXFP4 versions quantized via GPTQ.
  • GGUF: Various formats (Q8_0 through Q2_K) provided via llama.cpp.

Performance Results

Evaluation shows that reasoning capabilities scale with model size. The 30B model leads in agentic coding, achieving 57.00 on SWE Bench Verified and 29.24 on Terminal-Bench 2.1. Reasoning benchmarks such as AIME25 show a progression from 78.33 (3B) to 89.17 (30B). All models support 12 languages, including English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese.

Sources

Related

  • Dispatch
    IBM Granite 4.1 LLMs release notes / technical overview

    IBM has released Granite 4.1, a family of dense, decoder-only LLMs (3B, 8B, and 30B) that achieve high performance through rigorous data curation and a multi-stage reinforcement learning pipeline.

  • Dispatch
    IBM Granite 4.0 Nano Release Notes

    IBM has released Granite 4.0 Nano, a series of high-performance small language models (SLMs) designed for edge and on-device applications under an Apache 2.0 license.

  • Dispatch
    Granite 4.0 3B Vision release notes / what's new

    IBM has released Granite 4.0 3B Vision, a compact multimodal model optimized for enterprise document understanding, featuring high-accuracy table extraction, chart reasoning, and key-value pair extraction.

  • Dispatch
    Google Gemma 2 Release Notes

    Google has released Gemma 2, a family of open-weight LLMs available in 9B and 27B parameter sizes, featuring technical advances in distillation and attention mechanisms to improve performance.

  • Project
    zai-org/GLM-5

    GLM‑5 is an open‑source series of very large LLMs (320 B‑744 B total parameters) optimized for multimodal input, 1 M‑token contexts, and long‑horizon agentic tasks. It introduces hybrid sparse/linear attention, IndexShare, and a custom RL engine (slime) to achieve frontier‑level performance on coding and planning benchmarks while offering cheap FP8 inference. Models are downloadable from Hugging Face/ModelScope and can be served with SGLang, vLLM, KTransformers, etc., with optional ‘reasoning_effort’ control.