IBM Granite 4.2 8B

IBM Granite 4.2 8B

Quick answer: IBM Granite 4.2 8B is the mid-size model in IBM's reasoning-focused Granite 4.2 family — a dense, Apache 2.0 model that, unlike the 3B, completes IBM's full agentic-RL training block (SWE-agent, Terminal-agent, Search-agent). It scores 47.7% on SWE-bench Verified, 86.7% on AIME25, and 50.3% on BFCL-v4, with a native 128K-token context.

At a Glance

Where Granite 4.2 8B leads

• Full agentic-RL training (SWE-agent → Terminal-agent → Search-agent) — reports real scores on SWE-bench Multilingual (30.8%), SWE-bench Pro (19.1%), SWE-bench Verified (47.7%), and Terminal-Bench 2.1 (20.6%), unlike the 3B model
• 86.7% AIME25 and 78.3% HMMT Feb25 — strong competition-math reasoning for an 8B dense model
• Native tool calling in OpenAI function-calling format, integrates directly with OpenHands, OpenCode, and Pi harnesses

Where it lags

• Trails the 30B sibling on every reported metric — largest gaps on SWE-bench Pro (19.1% vs 33.3%) and GDPVal (1189 vs 1225 Elo)
• 20.6% Terminal-Bench 2.1 is modest for agentic terminal-operation reliability
• Not natively multimodal — text-only

Best for: Enterprise agentic coding/terminal/search workflows on mid-size hardware — teams wanting real agentic-RL-trained tool use without the 30B's compute footprint.

What Granite 4.2 8B Is

Granite 4.2 8B is the middle size 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 30B siblings, but — like the 30B — completes IBM's full post-training ladder: foundational RL (verifiable rewards + instruction-following/code skill boosters), then the agentic-RL block — three sequential GRPO stages trained on real, sandboxed environments: an OpenHands-driven SWE-agent stage (real repos, hidden-test verification), a Harbor/Terminus-2-driven Terminal-agent stage (live shell, up to 64 environment turns per rollout), and a browsing Search-agent stage (multi-hop web research, LLM-judged) — followed by final RLHF alignment.

Training data includes 1 trillion tokens of synthetic code generated via IBM's CodeAlchemy pipeline, plus a speculative-decoding layer for faster inference. Like all Granite 4.2 models, it supports a thinking/non-thinking/low-effort reasoning switch and native tool calling.

Specifications

Pricing

Open weights available on Hugging Face (ibm-granite/granite-4.2-8b) — 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 8B natively supports 131,072 tokens (128K).

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). Bird-SQL and GDPVal-AA v2 are reported by IBM under the shorthand names "BirdBench" (41.07%, converted to Benchgen's 0-1 fractional scale) and "GDPval" (1189, 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 (58.06) — 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 (41.20) — no Benchgen page exists.
• GPQA (64.14, plain — not "GPQA Diamond") — Benchgen only tracks the Diamond subset; skipped to avoid conflating distinct question sets.
• MMLU-ProX lite (IBM) (61.06) — explicitly labeled an IBM in-house variant.
• RULER 64K (80.99) and RULER 128K (71.41) — no Benchgen page exists for RULER yet.

Related Models & Benchmarks

• Compare: Granite 4.2 3B · Granite 4.2 30B · Granite 4.0 Tiny Preview
• Benchmarks: SWE-bench Verified · BFCL-v4 · AIME 2025 · TerminalBench

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.