Model Details
Granite 3.3 8B Base
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 | 89.7% | Benchgen evaluation | 2025-04 |
| HellaSwag | 80.1% | Benchgen evaluation | 2025-04 |
| GSM8K | 59.0% | Benchgen evaluation | 2025-04 |
| Arena Hard | 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 | 89.7% | — | Instruct | Apache 2.0 |
| 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 · Llama 3.1 8B Instruct
- Benchmarks: HumanEval · HellaSwag · 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.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.