Atlas

Models

← All models

Granite 3.3 8B Instruct

IBM · 2025-04-16 · 8.2B parameters

Granite 3.3 8B Instruct is IBM's instruction-tuned, text-only 8-billion-parameter model with a 131,072-token context window. It adds fill-in-the-middle generation and an optional thinking mode that emits structured reasoning in <think> and <response> tags; IBM reports improvements in mathematics, coding, and instruction following. Its weights are released under the Apache 2.0 license.

Benchmark scores

BenchmarkScore
AA-LCR4.3
AA-Omniscience Index-77.3
AIME 2025 (Artificial Analysis avg@10)6.7
Artificial Analysis Omniscience Accuracy9.7
Artificial Analysis Omniscience Hallucination Rate96.2
Artificial Analysis Progress Index v1131
Capability107.9
CritPt (avg@5)0.0
GPQA Diamond (Artificial Analysis avg@5)33.8
Humanity's Last Exam (Text-Only)4.2
IFBench (Artificial Analysis avg@5)22.4
LiveCodeBench (Artificial Analysis)12.7
MATH-500 (Artificial Analysis source)66.5
MMLU Pro46.8
SciCode (Artificial Analysis avg@3)10.1
Terminal-Bench Hard0.0
Vectara Hallucination Leaderboard10.6
τ²-bench (Telecom)10.5
Loading Atlas data…