IBM Granite 4.2 3B

IBM Granite 4.2 3B

Quick answer: IBM Granite 4.2 3B is the smallest model in IBM's reasoning-focused Granite 4.2 family — a dense, Apache 2.0 model with a thinking/non-thinking/low-effort switch. It scores 78.3% on AIME25, 67.8% on MMLU-Pro, and supports a native 128K-token context. Unlike its 8B/30B siblings, it does not receive the agentic-RL training stage (no SWE/Terminal/Search environments).

At a Glance

Where Granite 4.2 3B leads

• Strong reasoning-per-parameter: 78.3% AIME25 and 66.7% HMMT Feb25 at just 3B dense parameters
• Native step-by-step reasoning with a thinking / non-thinking / low-effort switch — spend more compute only when a task needs it
• Fully open under Apache 2.0 — no licensing restrictions on fine-tuning or production use

Where it lags

• No agentic-RL training — unlike the 8B/30B siblings, it wasn't trained on SWE-agent, terminal, or web-search environments, so it lacks reported scores on SWE-bench, Terminal-Bench, or GDPval
• Lower ceiling than its siblings across every reported benchmark (expected at this scale)
• 26.7-point gap vs. the 30B model on BFCL (v4) — weaker native tool-calling reliability

Best for: Lightweight enterprise reasoning/chat workloads (instruction following, math, general QA) where agentic tool-use isn't required — edge or high-throughput deployments where the 8B/30B agentic stack is overkill.

What Granite 4.2 3B Is

Granite 4.2 3B is the smallest of IBM's three-model Granite 4.2 release (3B/8B/30B), the first Granite generation built explicitly around reasoning. All three sizes share the same dense decoder-only transformer architecture and training pipeline — pre-trained from scratch on ~15 trillion tokens across a five-phase strategy, then supervised fine-tuned on chain-of-thought, reasoning, and agentic-trajectory data, then post-trained with a multi-stage GRPO reinforcement-learning pipeline.

Where the 3B model differs from its siblings is post-training depth: it completes the foundational RL stages (verifiable-reward RL plus instruction-following and code skill boosters) and final RLHF alignment, but takes a shortened path that skips the agentic RL block (SWE-agent, Terminal-agent, Search-agent environments) reserved for the 8B and 30B models. Every Granite 4.2 model — including the 3B — supports native tool calling in the OpenAI function-calling format and a low_effort thinking mode that spends a short reasoning budget on easy questions.

Specifications

Pricing

Open weights available on Hugging Face (ibm-granite/granite-4.2-3b) — free to self-host under Apache 2.0. FP8, NVFP4, MXFP4, and GGUF quantized variants are also published for reduced-memory deployment.

Context Window

Granite 4.2 3B natively supports 131,072 tokens (128K). IBM's five-phase pre-training pipeline includes a dedicated long-context phase.

Public Benchmark Scores

Scores are self-reported by IBM Research on the official Hugging Face launch blog (huggingface.co/blog/ibm-granite/granite-4-2), which reports results for all three Granite 4.2 sizes in one shared table. IFBench is reported by IBM as "IFBench (prompt)" — the prompt-level-loose submetric, matching Benchgen's ifbench metric exactly.

Benchmarks reported for 8B/30B but not applicable to the 3B model (no agentic-RL training): SWE-bench Multilingual, SWE-bench Pro, SWE-bench Verified, Terminal-Bench 2.1, Bird-SQL (BirdBench), GDPVal — all reported as "NA" for the 3B size in IBM's own table.

Additional benchmarks reported by IBM, not added to Benchgen:

• τ³-bench (45.78) — IBM's table doesn't specify a domain, and Benchgen's only τ³ page (tau3-banking) covers a much lower-ceiling banking-specific split (max observed score 33.4 vs. Granite's 45.78) — scale mismatch indicates these aren't the same evaluation; skipped rather than guessed.
• ProfBench (32.10) — no Benchgen page exists for this benchmark.
• GPQA (54.80, plain — not "GPQA Diamond") — Benchgen only tracks the Diamond subset; the main GPQA question set is a distinct benchmark and conflating the two risks a wrong-scale/wrong-question-set entry.
• MMLU-ProX lite (IBM) (27.78) — explicitly labeled an IBM in-house variant, not independently reproducible.
• RULER 64K (67.52) and RULER 128K (55.30) — no Benchgen page exists for RULER yet.

Related Models & Benchmarks

• Compare: Granite 4.2 8B · Granite 4.2 30B · Granite 4.0 Tiny Preview
• Benchmarks: AIME 2025 · MMLU-Pro · LiveCodeBench v6 · IFBench

Frequently Asked Questions

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.