Granite 4.2 8B

// Overview

Granite 4.2 8B is the mid-size model of the three IBM published on 25 August 2026 as its first family of dense, decoder-only reasoning language models. Like its siblings it ships with an explicit thinking mode — chain-of-thought inside <think>…</think> tags before the answer — which callers can switch off, or run at low effort, to trade reasoning depth for latency. The weights are on Hugging Face under the Apache 2.0 license, with builds also distributed through Ollama and GitHub.

It shares the family recipe: pre-training from scratch on roughly 15 trillion tokens across a five-phase schedule, supervised fine-tuning on about 7.2 million chain-of-thought, reasoning and agentic-trajectory samples, then a multi-stage GRPO reinforcement-learning pipeline. Along with the 30B — and unlike the 3B — the 8B also goes through agentic RL inside real software-engineering, terminal and web-search environments, which is why it carries scores on the agentic coding benchmarks the 3B leaves blank. Context is 128K tokens natively with a documented extension to 512K, tool calling uses the OpenAI-compatible function format, and twelve languages are supported.

On IBM's published evaluations the 8B scores 86.67 on AIME25, 64.14 on GPQA, 73.24 on LiveCodeBench v6, 47.67 on SWE-bench Verified, 74.04 on MMLU-Pro and 71.41 on RULER at 128K, and it is the only size to beat the 30B on a row — IFBench prompt-following, 79.33 against 77.17. It runs under Transformers, vLLM or SGLang, with FP8, NVFP4, MXFP4 and GGUF quantizations published alongside the bf16 weights.

// Benchmarks

This model's scores

1. AIME2586.67%
2. RULER 64K80.99%
3. IFBench (prompt)79.33%
4. HMMT Feb2578.33%
5. MMLU-Pro74.04%
6. LiveCodeBench v673.24%
7. RULER 128K71.41%
8. Arena-Hard-V265.19%
9. GPQA64.14%
10. τ³-bench58.06%
11. BFCL v450.29%
12. SWE-bench Verified47.67%
13. ProfBench41.2%
14. BirdBench41.07%
15. SciCode36.09%
16. SWE-bench Multilingual30.78%
17. Terminal-Bench 2.120.56%
18. SWE-bench Pro19.11%

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 on software-engineering, terminal and search environments, which the 3B skips
• Best IFBench prompt-following score in the family at 79.33, ahead of the 30B
• 128K native context with 71.41 on RULER at 128K
• FP8, NVFP4, MXFP4 and GGUF builds published alongside the bfloat16 weights

// Best for

• Single-GPU coding and tool-calling agents that need permissively-licensed weights
• On-premises enterprise assistants where inference cost matters more than peak accuracy
• Instruction-following workloads that reward the family's strongest IFBench score
• Fine-tuning a small Apache-2.0 reasoning base on domain-specific tasks

// 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.