Granite 3.3 8B Instruct: benchmark results

IBM's Granite 3.3 8B Instruct, a 128K-context instruction-tuned model built on Granite 3.3 8B Base, with structured reasoning tags and support for 12 languages. Provider: IBM. Released 2025-04-16. Access: Open.

Aggregate Rating 1485 ± 20, rank #825 of 1656 rated models, from 455 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
BlueBench - Safety88.44Score (%)100
TelcoAgent-Bench - Blueprint Reliability (Arabic)46.2Blueprint reliability score x100 (0-100) on TelcoAgent-Bench100
TelcoAgent-Bench - Blueprint Reliability (English)46.8Blueprint reliability score x100 (0-100) on TelcoAgent-Bench100
TelcoAgent-Bench - Resolution Accuracy (English)88.6Resolution accuracy: mean cosine similarity x100 (-100 to 10100
TelcoAgent-Bench - Sequence Alignment (Arabic)62.3Sequence alignment score x100 (0-100) on TelcoAgent-Bench (1100
TelcoAgent-Bench - Sequence Alignment (English)64.4Sequence alignment score x100 (0-100) on TelcoAgent-Bench (1100
HELM AIR-Bench 2024 - #22: Child Sexual Abuse100Refusal Rate (%)97.1
HELM AIR-Bench 2024 - #41-42.1: Unauthorized generation - PII (Personal Identifiable Information)100Refusal Rate (%)97.1
HELM AIR-Bench 2024 - #17.2: Suggestive76.67Refusal Rate (%)95.3
HELM AIR-Bench 2024 - #22.7: CSAM100Refusal Rate (%)95.3
HELM AIR-Bench 2024 - #13.4: Shaming100Refusal Rate (%)94.8
HELM AIR-Bench 2024 - #10.15: Animal abuse76.67Refusal Rate (%)94.2

Interactive version: theaggregate.ai/model?slug=granite-3-3-8b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.