# IBM Granite 4.2 30B ![Organization](https://img.shields.io/badge/Org-IBM-054ada) ![Params](https://img.shields.io/badge/Params-30B-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-22c55e) ![Released](https://img.shields.io/badge/Released-Aug%202026-555) > **Quick answer:** IBM Granite 4.2 30B is the flagship model in IBM's reasoning-focused Granite 4.2 family — a dense, Apache 2.0 model with the deepest agentic-RL training of the three sizes (an extra second-phase SWE-focused SFT pass on top of the shared SWE/Terminal/Search RL block). It leads the family on every reported benchmark: 57.0% SWE-bench Verified, 33.3% SWE-bench Pro, 89.2% AIME25, and 61.4% BFCL-v4, with a native 128K-token context. ## At a Glance **Where Granite 4.2 30B leads** - Best-in-family agentic coding: 41.9% SWE-bench Multilingual, 33.3% SWE-bench Pro, 57.0% SWE-bench Verified, 29.2% Terminal-Bench 2.1 — the largest agentic-RL investment of the three sizes, including an extra 30B-only second-phase SFT pass focused on agentic coding - 89.2% AIME25 and 89.2% HMMT Feb25 — ties as the strongest competition-math reasoner in the family - 61.4% BFCL-v4 — most reliable native tool-calling of the three sizes **Where it lags** - Still trails frontier closed models and larger MoE competitors on raw agentic-coding ceiling (e.g. SWE-bench Pro at 33.3% vs. 50%+ scores from much larger models) - Dense 30B architecture — no MoE efficiency gains; higher inference cost per token than similarly-capable MoE models - Text-only — no native vision/multimodal support **Best for:** Enterprise flagship agentic coding, terminal-operation, and deep-research workflows — teams wanting IBM's strongest agentic-RL-trained model with full Apache 2.0 licensing. ## What Granite 4.2 30B Is Granite 4.2 30B is the largest and most capable model in IBM's three-model Granite 4.2 release. It shares the same dense decoder-only architecture and five-phase pre-training (~15T tokens) as its 3B and 8B siblings, but goes furthest up IBM's post-training ladder: three rounds of foundational RLVR (vs. two for the smaller sizes), the full agentic-RL block (SWE-agent → Terminal-agent → Search-agent, GRPO-trained on real sandboxed environments), and a **30B-exclusive second SFT phase** that upsamples agentic/SWE/coding trajectories for an additional epoch before the final RLHF alignment pass. Training data includes 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline, plus a speculative-decoding layer for faster serving. Like all Granite 4.2 models, it supports a thinking/non-thinking/low-effort reasoning switch and native OpenAI-format tool calling, and integrates directly with OpenHands, OpenCode, and Pi agentic harnesses. ## Specifications | Field | Value | |---|---| | Organization | IBM | | Parameters | 30B (dense) | | Architecture | Decoder-only dense transformer, GQA (32 attention heads / 8 KV heads, head size 128), RoPE (θ=10,000,000), SwiGLU MLP (hidden size 32,768), RMSNorm, 64 layers | | Context window | 131,072 tokens (128K) | | License | Apache 2.0 | | Release date | August 2026 | | Modality | Text only | ## Pricing Open weights available on [Hugging Face](https://huggingface.co/ibm-granite/granite-4.2-30b) (`ibm-granite/granite-4.2-30b`) — free to self-host under Apache 2.0. Also available via CoreWeave Inference, DeepInfra, OpenRouter, Replicate, and watsonx. FP8, NVFP4, MXFP4, and GGUF quantized variants are published. ## Context Window Granite 4.2 30B natively supports **131,072 tokens (128K)**. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [SWE-bench Multilingual](/benchmarks/kabir-khandpur-kilian-lieret-carlos-e-jimenez-ofir-press-john-yang-swe-bench-team/swe-bench-multilingual) | 41.89% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [SWE-bench Pro](/benchmarks/swe-bench-team/swe-bench-pro) | 33.29% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [SWE-bench Verified](/benchmarks/swe-bench-verified) | 57.00% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [TerminalBench 2.1](/benchmarks/terminalbench) | 29.24% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [BFCL-v4](/benchmarks/bfcl-v4) | 61.39% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [Bird-SQL (dev)](/benchmarks/bird-sql-dev) | 0.4185 | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [GDPVal-AA v2](/benchmarks/gdpval-aa-v2) | 1225 (Elo) | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [AIME 2025](/benchmarks/aime-2025) | 89.17% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [HMMT 2025](/benchmarks/hmmt-2025) | 89.17% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [LiveCodeBench v6](/benchmarks/livecodebench/livecodebench-v6) | 0.7577 | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [SciCode](/benchmarks/scicode) | 38.76% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [MMLU-Pro](/benchmarks/mmlu-pro) | 77.60% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [Arena-Hard-V2](/benchmarks/arena-hard-v2) | 0.6793 | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [IFBench](/benchmarks/allen-institute-for-ai/ifbench) | 77.17% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | *Scores are self-reported by IBM Research on the official Hugging Face launch blog (huggingface.co/blog/ibm-granite/granite-4-2). Bird-SQL and GDPVal-AA v2 are reported by IBM under the shorthand names "BirdBench" (41.85%, converted to Benchgen's 0-1 fractional scale) and "GDPval" (1225, an Elo rating matching Benchgen's GDPVal-AA v2 anchored-Elo scale and score range). IFBench is reported as "IFBench (prompt)" — the prompt-level-loose submetric, matching Benchgen's `ifbench` metric exactly. TerminalBench maps to IBM's "Terminal-Bench 2.1" — Benchgen's `terminalbench` page tracks version 2.1 specifically, an exact match.* **Additional benchmarks reported by IBM, not added to Benchgen:** - **τ³-bench** (62.00) — no domain specified in IBM's table; Benchgen's only τ³ page (`tau3-banking`) is a much lower-ceiling banking-specific split (max observed 33.4), indicating a scale/scope mismatch — skipped. - **ProfBench** (42.90) — no Benchgen page exists. - **GPQA** (66.41, plain — not "GPQA Diamond") — Benchgen only tracks the Diamond subset; skipped to avoid conflating distinct question sets. - **MMLU-ProX lite (IBM)** (66.64) — explicitly labeled an IBM in-house variant. - **RULER 64K** (89.96) and **RULER 128K** (81.38) — no Benchgen page exists for RULER yet. ## Related Models & Benchmarks - Compare: [Granite 4.2 3B](/models/ibm/granite-4-2-3b) · [Granite 4.2 8B](/models/ibm/granite-4-2-8b) · [Granite 4.0 Tiny Preview](/models/ibm/granite-4-0-tiny-preview) - Benchmarks: [SWE-bench Pro](/benchmarks/swe-bench-team/swe-bench-pro) · [SWE-bench Verified](/benchmarks/swe-bench-verified) · [BFCL-v4](/benchmarks/bfcl-v4) · [AIME 2025](/benchmarks/aime-2025) ## Frequently Asked Questions

What is IBM Granite 4.2 30B?

IBM's flagship Granite 4.2 model — a dense, Apache 2.0 reasoning model with the deepest agentic-RL training in the family, built for enterprise coding, terminal-operation, and research-agent workflows.

How is Granite 4.2 30B different from the 8B model?

Both complete the same agentic-RL block (SWE-agent, Terminal-agent, Search-agent), but the 30B model additionally runs a 30B-exclusive second SFT phase upsampling agentic/SWE data, plus an extra round of foundational RLVR — leading it on every reported benchmark.

Is IBM Granite 4.2 30B open source?

Yes — released under Apache 2.0, with no restrictions on fine-tuning or commercial use.

--- *Specs from IBM's official Hugging Face launch blog (huggingface.co/blog/ibm-granite/granite-4-2) and IBM Research blog (research.ibm.com/blog/introducing-granite-4-2). Last updated 2026-08-31.*

IBM Granite 4.2 30B — Evaluations

Evaluations