# Granite 3.3 8B Base ![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) ![Type](https://img.shields.io/badge/Type-Base%20Model-gray) > **Quick answer:** IBM Granite 3.3 8B Base is the April 2025 pre-trained base model underlying Granite 3.3 8B Instruct, scoring 89.7% HumanEval, 80.1% HellaSwag, 59.0% GSM8K, and 57.6% Arena Hard. Apache 2.0. ## At a Glance **Where Granite 3.3 8B Base leads** - 89.7% HumanEval — strong coding capability from base - 80.1% HellaSwag — solid language understanding - Apache 2.0 — fully open for fine-tuning - Good foundation for custom fine-tuning **Where it lags** - 59.0% GSM8K — moderate math for a base model - Base model: not instruction-tuned; use Instruct variant for deployment - Requires fine-tuning or prompting for production use **Best for:** Fine-tuning base for domain-specific models; research and experimentation; teams building custom instruction-tuned variants. ## What Granite 3.3 8B Base Is Granite 3.3 8B Base is the pre-trained (non-instruction-tuned) foundation model from IBM's Granite 3.3 series. Released April 2025 alongside Granite 3.3 8B Instruct, it is intended as a starting point for fine-tuning on domain-specific tasks. IBM publishes base models alongside instruct variants to support the research and fine-tuning community. The identical HumanEval (89.7%) and Arena Hard (57.6%) scores between Base and Instruct suggest the benchmark tasks benefit from the base model's pre-training rather than instruction tuning specifically. ## Specifications | Field | Value | |---|---| | Organization | IBM | | License | Apache 2.0 | | HuggingFace | ibm-granite/granite-3.3-8b-base | | Release date | April 2025 | | Parameters | 8B | | Type | Base (pre-trained) | | Modality | Text only | ## Pricing Open weights under Apache 2.0 — self-host at no cost. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [HumanEval](/benchmarks/humaneval) | 89.7% | Benchgen evaluation | 2025-04 | | [HellaSwag](/benchmarks/hellaswag) | 80.1% | Benchgen evaluation | 2025-04 | | [GSM8K](/benchmarks/gsm8k) | 59.0% | Benchgen evaluation | 2025-04 | | [Arena Hard](/benchmarks/arena-hard-v2) | 57.6% | Benchgen evaluation | 2025-04 | ## Granite 3.3 8B Base vs Alternatives | Model | HumanEval | HellaSwag | Type | License | |---|---|---|---|---| | **Granite 3.3 8B Base** | 89.7% | 80.1% | Base | Apache 2.0 | | [Granite 3.3 8B Instruct](/models/ibm/granite-3-3-8b-instruct) | 89.7% | — | Instruct | Apache 2.0 | | [Llama 3.1 8B Instruct](/models/meta/llama-3-1-8b-instruct) | — | — | Instruct | Llama 3.1 | For most deployments, use Granite 3.3 8B Instruct (instruction-tuned). Use Granite 3.3 8B Base when fine-tuning for a specific domain or task. ## Related Models & Benchmarks - Compare: [Granite 3.3 8B Instruct](/models/ibm/granite-3-3-8b-instruct) · [Llama 3.1 8B Instruct](/models/meta/llama-3-1-8b-instruct) - Benchmarks: [HumanEval](/benchmarks/humaneval) · [HellaSwag](/benchmarks/hellaswag) · [GSM8K](/benchmarks/gsm8k) ## Frequently Asked Questions

What is Granite 3.3 8B Base?

IBM's April 2025 pre-trained base model (non-instruction-tuned) scoring 89.7% HumanEval and 80.1% HellaSwag. Apache 2.0 — use as a fine-tuning foundation or use Instruct variant for direct deployment.

Should I use Granite 3.3 8B Base or Granite 3.3 8B Instruct?

Granite 3.3 8B Instruct for direct deployment and chat/instruction use. Granite 3.3 8B Base for fine-tuning on domain-specific data.

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

Granite 3.3 8B Base — Training jobs

Training jobs