# Granite 3.3 8B Instruct ![Organization](https://img.shields.io/badge/Org-IBM-1f70c1) ![License](https://img.shields.io/badge/License-Apache%202.0-22c55e) ![Released](https://img.shields.io/badge/Released-Apr%202025-555) ![Params](https://img.shields.io/badge/Params-8B-blue) > **Quick answer:** IBM Granite 3.3 8B Instruct is IBM's April 2025 compact model scoring 89.7% HumanEval, 80.9% GSM8K, and 57.6% Arena Hard. Apache 2.0 licensed with strong coding performance for an 8B model. ## At a Glance **Where Granite 3.3 8B Instruct leads** - 89.7% HumanEval — exceptional coding for an 8B model (matches many larger models) - Apache 2.0 — fully open including commercial use - IBM enterprise trust: built for regulated enterprise environments - Competitive at its parameter count **Where it lags** - 57.6% Arena Hard — moderate on open-ended instruction quality - 80.9% GSM8K — good but below leading 8B models - 8B parameters: limited capacity vs larger models - Smaller context window than MoE competitors **Best for:** Enterprise coding assistants; IBM Cloud deployments; regulated industries requiring Apache 2.0 small models with strong HumanEval scores. ## What Granite 3.3 8B Instruct Is Granite 3.3 8B Instruct is IBM's April 2025 update to the Granite 3 model family. IBM's Granite models are designed with enterprise use cases in mind: transparency (IBM publishes training data details and model cards), safety filtering, and compliance with enterprise data policies. The 89.7% HumanEval score is exceptional for an 8B model — placing it above many larger open-source models from 2024. This reflects IBM's focus on coding capability in its developer tooling (Granite was built for tasks like code completion, bug fixing, and documentation). ## Specifications | Field | Value | |---|---| | Organization | IBM | | License | Apache 2.0 | | HuggingFace | ibm-granite/granite-3.3-8b-instruct | | Release date | April 2025 | | Parameters | 8B | | Modality | Text only | ## Pricing Open weights under Apache 2.0 — self-host at no license cost. Available via IBM watsonx.ai and cloud providers. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [HumanEval](/benchmarks/humaneval) | 89.7% | Benchgen evaluation | 2025-04 | | [GSM8K](/benchmarks/gsm8k) | 80.9% | Benchgen evaluation | 2025-04 | | [Arena Hard](/benchmarks/arena-hard-v2) | 57.6% | Benchgen evaluation | 2025-04 | ## Granite 3.3 8B Instruct vs Alternatives | Model | HumanEval | GSM8K | Params | License | |---|---|---|---|---| | **Granite 3.3 8B Instruct** | 89.7% | 80.9% | 8B | Apache 2.0 | | [Llama 3.1 8B Instruct](/models/meta/llama-3-1-8b-instruct) | — | 84.5% | 8B | Llama 3.1 | | [Gemma 3 12B](/models/google/gemma-3-12b) | 85.4% | — | 12B | Gemma ToU | | [Phi-4](/models/microsoft/phi-4) | — | — | 14B | MIT | Granite 3.3 8B Instruct leads on HumanEval (89.7%) among compact open-weight models. For pure coding in enterprise/IBM environments: Granite 3.3 8B. For broader instruction quality: Llama 3.1 8B Instruct. ## Related Models & Benchmarks - Compare: [Llama 3.1 8B Instruct](/models/meta/llama-3-1-8b-instruct) · [Gemma 3 12B](/models/google/gemma-3-12b) · [Phi-4](/models/microsoft/phi-4) - Benchmarks: [HumanEval](/benchmarks/humaneval) · [GSM8K](/benchmarks/gsm8k) · [Arena Hard](/benchmarks/arena-hard-v2) ## Frequently Asked Questions

What is Granite 3.3 8B Instruct?

IBM's April 2025 compact 8B model scoring 89.7% HumanEval, 80.9% GSM8K, and 57.6% Arena Hard. Apache 2.0 with enterprise-focused design.

Why is Granite 3.3's HumanEval so high for an 8B model?

IBM specifically optimised Granite for code tasks — it is a primary use case for IBM's developer tooling (code completion, documentation, bug fixing). The 89.7% HumanEval reflects this coding-first approach.

--- *Specs from IBM's Granite 3.3 8B Instruct release (April 2025) and Benchgen evaluations. Last updated 2026-07-24.*

Granite 3.3 8B Instruct — Evaluations

Evaluations