Granite 4.2 30B

// Overview

Granite 4.2 30B is the largest of the three models IBM published on 25 August 2026 as its first family of dense, decoder-only reasoning language models. Where earlier Granite releases were instruction-tuned chat models, every Granite 4.2 model ships with an explicit thinking mode: the model emits chain-of-thought inside <think>…</think> tags before answering, and callers can switch that reasoning off, or run it at low effort, to trade depth for latency. The weights are on Hugging Face under the Apache 2.0 license, with builds also distributed through Ollama and GitHub.

The training recipe starts with pre-training from scratch on roughly 15 trillion tokens across a five-phase schedule that extends the context window, followed by supervised fine-tuning on about 7.2 million samples of chain-of-thought, reasoning and agentic-trajectory data. Post-training is a multi-stage GRPO reinforcement-learning pipeline; the 8B and 30B models additionally go through agentic RL inside real software-engineering, terminal and web-search environments, which is the step the 3B model skips. Context is 128K tokens natively with a documented long-context extension to 512K, and the model supports native OpenAI-compatible tool calling across twelve languages including English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese.

On IBM's published evaluations the 30B leads its two smaller siblings across the board: 89.17 on AIME25, 66.41 on GPQA, 75.77 on LiveCodeBench v6, 57.00 on SWE-bench Verified, 29.24 on Terminal-Bench 2.1, 77.60 on MMLU-Pro and 81.38 on RULER at 128K. It runs under Transformers, vLLM or SGLang, and IBM publishes FP8, NVFP4, MXFP4 and GGUF quantizations alongside the bf16 weights.

// Benchmarks

This model's scores

1. AIME2589.17%
2. HMMT Feb2589.17%
3. RULER 64K89.96%
4. RULER 128K81.38%
5. MMLU-Pro77.6%
6. LiveCodeBench v675.77%
7. Arena-Hard-V267.93%
8. GPQA66.41%
9. τ³-bench62%
10. BFCL v461.39%
11. SWE-bench Verified57%
12. IFBench (prompt)77.17%
13. ProfBench42.9%
14. SWE-bench Multilingual41.89%
15. BirdBench41.85%
16. SciCode38.76%
17. SWE-bench Pro33.29%
18. Terminal-Bench 2.129.24%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Apache 2.0 weights — no licence restrictions on commercial use, fine-tuning or self-hosting
• Switchable thinking mode: full chain-of-thought, low-effort reasoning, or direct answers from one checkpoint
• Agentic reinforcement learning inside real software-engineering, terminal and search environments
• Strongest long-context scores in the family — RULER 89.96 at 64K and 81.38 at 128K
• Native OpenAI-compatible tool calling and 12-language support
• FP8, NVFP4, MXFP4 and GGUF builds published alongside the bfloat16 weights

// Best for

• Self-hosted coding and terminal agents that need permissively-licensed weights
• Enterprise deployments where data cannot leave on-premises or private-cloud infrastructure
• Long-document analysis and retrieval workloads using the 128K native context
• Fine-tuning an Apache-2.0 reasoning base on domain-specific tool use

// Granite (open-weight) — every version

The full lineage of the Granite (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.