Granite 4.2

model
IBM

IBM's first family of native reasoning Granite models: dense decoder-only Transformers in 3B, 8B and 30B, each with a built-in <think> chain-of-thought that can be toggled per request (full thinking, non-thinking, or a "low-effort" short budget) plus reasoning-augmented native tool calling. All three are post-trained on the Granite 4.1 base checkpoints (the model cards declare base_model: granite-4.1-{3b,8b,30b}-base; the 4.1 bases were pretrained on ~15T tokens with a five-phase curriculum) rather than being a fresh pretrain. Same GQA/RoPE/SwiGLU/RMSNorm shapes as 4.1 (30B: 64 layers, hidden 4096, MLP 32768, 32 heads / 8 KV heads, 100K vocab), 128K native context with 512K long-context extension. Apache 2.0.

The new work is the post-training pipeline. SFT uses ~7.2M samples (~100B tokens), 31.6% agentic — mostly software-engineering trajectories collected through OpenHands, OpenCode, Terminus-2, SWE-agent, Codex, Gemini CLI, Goose and other harnesses — with the 30B getting a second agentic-coding SFT phase. Then a staged, asynchronous GRPO curriculum on NeMo RL / NeMo Gym: RLVR and skill boosters for all sizes; for 8B and 30B an agentic RL block that runs SWE (OpenHands, hidden-test rewards) → terminal (Harbor/Terminus-2, up to 64 environment turns) → deep-research search agents in real sandboxes; and a closing RLHF phase. IBM also cites 1T tokens of CodeAlchemy synthetic code and a mid-training step. Trained on a CoreWeave GB200 NVL72 cluster. FP8, NVFP4, MXFP4 (LLM Compressor) and GGUF quants ship for every size.

30B results: SWE-bench Verified 57.0, SWE-bench Pro 33.3, Terminal-Bench 2.1 29.2, τ³-bench 62.0, BFCL v4 61.4, AIME25 89.2, GPQA 66.4, LiveCodeBench v6 75.8, MMLU-Pro 77.6, RULER-128K 81.4. AA Intelligence Index v4.3: 15 (30B), 11 (8B), 9 (3B) — versus 7 / 7 / 6 for the non-reasoning Granite 4.1 instruct models on the same bases. Released alongside Granite Speech 5.0 Turbo CTC (470M, no LLM backbone, ~12,600 RTFx on one H200).

IBM Research BlogHuggingFace Blog (How They're Built)GitHubIBM Granite DocsHuggingFace CollectionHuggingFace (30B)HuggingFace (8B)HuggingFace (3B)HuggingFace (30B GGUF)Artificial AnalysisAA (8B)AA (3B)

Model Details

Architecture DENSE
Parameters 30B
Context window 131,072
AA Intelligence 15 was 17 on v4.2
Training hardware NVIDIA GB200 NVL72 cluster (CoreWeave)
Optimizer Asynchronous GRPO (NeMo RL)
License Apache 2.0
Base model granite-4.1

Benchmark Scores

Benchmark Score Mode
SWE-bench Verified 57.0 30B
SWE-bench Pro 33.29 30B
SWE-bench Multilingual 41.89 30B
Terminal-Bench 2.1 29.24 30B
τ³-bench 62.00 avg (30B)
BFCL v4 61.39 30B
AIME25 89.17 30B
HMMT Feb25 89.17 30B
GPQA 66.41 30B
LiveCodeBench v6 75.77 30B
SciCode 38.76 30B
MMLU-Pro 77.60 30B
Arena-Hard-V2 67.93 30B
IFBench 77.17 prompt (30B)
RULER 128K 81.38 30B

Variants

Name Parameters Notes
Granite 4.2 30B 30B Post-trained on Granite-4.1-30B-Base; full agentic-RL ladder plus a second agentic-coding SFT phase. AA Intelligence Index v4.3 score 15 — highest in family.
Granite 4.2 8B 8B Post-trained on Granite-4.1-8B-Base; full agentic-RL ladder (SWE → terminal → search). AA Intelligence Index v4.3 score 11.
Granite 4.2 3B 3B Post-trained on Granite-4.1-3B-Base; foundational RL + RLHF only (no agentic-RL block). AA Intelligence Index v4.3 score 9.
Granite 4.2 FP8 / NVFP4 / MXFP4 / GGUF — vLLM quants for all three sizes (FP8 dynamic per-channel; FP4 via GPTQ on 2K SFT samples) and llama.cpp GGUF Q2_K–Q8_0.
open-weightenterprisereasoningagentic

Related