# IBM Granite 4.2 8B ![Organization](https://img.shields.io/badge/Org-IBM-054ada) ![Params](https://img.shields.io/badge/Params-8B-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 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 | Field | Value | |---|---| | Organization | IBM | | Parameters | 8B (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 12,800), RMSNorm, 40 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-4.2-8b) (`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 | 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) | 30.78% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [SWE-bench Pro](/benchmarks/swe-bench-team/swe-bench-pro) | 19.11% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [SWE-bench Verified](/benchmarks/swe-bench-verified) | 47.67% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [TerminalBench 2.1](/benchmarks/terminalbench) | 20.56% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [BFCL-v4](/benchmarks/bfcl-v4) | 50.29% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [Bird-SQL (dev)](/benchmarks/bird-sql-dev) | 0.4107 | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [GDPVal-AA v2](/benchmarks/gdpval-aa-v2) | 1189 (Elo) | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [AIME 2025](/benchmarks/aime-2025) | 86.67% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [HMMT 2025](/benchmarks/hmmt-2025) | 78.33% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [LiveCodeBench v6](/benchmarks/livecodebench/livecodebench-v6) | 0.7324 | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [SciCode](/benchmarks/scicode) | 36.09% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [MMLU-Pro](/benchmarks/mmlu-pro) | 74.04% | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [Arena-Hard-V2](/benchmarks/arena-hard-v2) | 0.6519 | Granite 4.2 launch blog (Hugging Face) | 2026-08 | | [IFBench](/benchmarks/allen-institute-for-ai/ifbench) | 79.33% | 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" (41.07%, 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](/models/ibm/granite-4-2-3b) · [Granite 4.2 30B](/models/ibm/granite-4-2-30b) · [Granite 4.0 Tiny Preview](/models/ibm/granite-4-0-tiny-preview) - Benchmarks: [SWE-bench Verified](/benchmarks/swe-bench-verified) · [BFCL-v4](/benchmarks/bfcl-v4) · [AIME 2025](/benchmarks/aime-2025) · [TerminalBench](/benchmarks/terminalbench) ## Frequently Asked Questions

What is IBM Granite 4.2 8B?

The mid-size model in IBM's Granite 4.2 family — a dense, Apache 2.0 reasoning model with full agentic-RL training (SWE-agent, Terminal-agent, Search-agent), built for enterprise coding, terminal, and research agent workflows.

Does Granite 4.2 8B support agentic coding?

Yes — it's trained through IBM's agentic-RL block on real sandboxed SWE, terminal, and web-search environments, and reports scores on SWE-bench Multilingual, SWE-bench Pro, SWE-bench Verified, and Terminal-Bench 2.1.

Is IBM Granite 4.2 8B 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 8B — Evaluations

Evaluations