Gemma 2B — Benchmarks

Benchmark scores for Gemma 2B aggregated from public leaderboards (10 boards), with how it ranks among open models and how it compares with its nearest neighbours. See hardware requirements for what you need to run it.

No llmrun Composite Score for Gemma 2B: the score is only published for a model measured on at least 4 of the boards in the score panel, spanning at least 2 benchmark categories and including a reasoning or coding board — or on 3 boards that narrow its score to a 90% range of at most 15.6 points. Either way the results must come from at least two different benchmarks. Its individual scores below stand on their own — see how the llmrun Composite Score gets calculated.

Gemma 2B reasoning benchmarks

Multi-step logic and problem-solving — can the model think through a hard problem rather than recall an answer.

BIG-Bench Hard
BIG-Bench Hard (BBH) is the 23-task subset of BIG-Bench that earlier models found hardest — multi-step reasoning, logic, and word problems. It's a long-standing standard for tracking reasoning ability across model generations.
35.2%
35.2%
median 50.8%open leader 87.5%
Overall best: Gemini 1.5 Pro 001 89.2%
16th percentile · #32 of 38 open · #44 of 50 all
#32 of 38
#44 of 50
Lab-reported
Epoch AI · Feb 2024

Other results (2)· leaderboard-only, not part of the llmrun Score

WinoGrande
WinoGrande is a large-scale commonsense-reasoning test built around pronoun resolution: the model must pick which of two nouns a pronoun refers to in a sentence, using real-world knowledge to disambiguate. It was designed by AI2 to resist the shortcuts that let earlier models game its predecessor, Winograd.
65.4%
65.4%
median 70.1%open leader 89.2%
36th percentile · #33 of 51 open · #60 of 79 all
#33 of 51
#60 of 79
Lab-reported
Epoch AI · Feb 2024
PIQA
PIQA (Physical Interaction QA) asks a model to choose the more sensible of two ways to accomplish an everyday physical task, testing practical, physical-world commonsense that isn't well captured by purely textual reasoning tests.
77.3%
77.3%
median 80.1%open leader 88.6%
Overall best: GPT 4o Mini (Jul 18, 2024) 88.7%
33rd percentile · #27 of 40 open · #45 of 59 all
#27 of 40
#45 of 59
Lab-reported
Epoch AI · Feb 2024

Gemma 2B math benchmarks

Symbolic and quantitative problem-solving, from competition math to multi-step calculation.

GSM8K
GSM8K is a classic set of grade-school math word problems requiring several arithmetic reasoning steps. Long a default sanity check for reasoning, top models now score near the ceiling, so it best separates smaller models.
17.7%
17.7%
median 52.3%open leader 94.5%
11th percentile · #55 of 62 open · #75 of 93 all
#55 of 62
#75 of 93
Lab-reported
Epoch AI · Feb 2024

Gemma 2B knowledge benchmarks

Breadth and depth of factual knowledge across many subjects, paired with the reasoning to apply it.

MMLU-Pro
MMLU-Pro is a harder, cleaned-up successor to MMLU with ten answer choices and more reasoning-heavy questions across 14 subjects, measuring broad knowledge and reasoning together.
15.8%
15.8%
median 52.1%open leader 88.0%
Overall best: Gemini-3.1-Pro 91.2%
8th percentile · #134 of 145 open · #245 of 257 all
#134 of 145
#245 of 257
Independent
TIGER-Lab

Other results (5)· leaderboard-only, not part of the llmrun Score

MMLU
MMLU (Massive Multitask Language Understanding) spans 57 subjects from history to law to medicine as multiple-choice questions. It is the long-standing default for broad knowledge and remains the most widely-reported general benchmark.
42.3%
42.3%
median 62.5%open leader 87.2%
Overall best: GPT 4o (Nov 20, 2024) 88.1%
21st percentile · #68 of 85 open · #117 of 136 all
#68 of 85
#117 of 136
Lab-reported
Epoch AI · Feb 2024
ARC Challenge
ARC Challenge is a set of grade-school science questions, curated by the Allen Institute for AI (AI2) to exclude ones that simple word-matching or retrieval can answer. It measures science knowledge paired with the reasoning needed to apply it.
42.1%
42.1%
median 53.9%open leader 95.3%
31st percentile · #39 of 56 open · #58 of 76 all
#39 of 56
#58 of 76
Lab-reported
Epoch AI · Feb 2024
HellaSwag
HellaSwag is a commonsense benchmark that asks a model to pick the most plausible continuation of an everyday situation. The wrong options are adversarially chosen to fool models, testing grounded commonsense understanding.
71.4%
71.4%
median 76.2%open leader 89.2%
Overall best: GPT 4 (Mar 14) 95.3%
37th percentile · #30 of 47 open · #50 of 75 all
#30 of 47
#50 of 75
Lab-reported
Epoch AI · Feb 2024
BoolQ
BoolQ is a reading-comprehension benchmark of naturally occurring yes/no questions, each paired with a short passage that contains the answer. It tests whether a model can extract and reason over information from real text rather than just recall facts.
69.4%
69.4%
median 76.4%open leader 89.4%
Overall best: T5 11B 91.2%
32nd percentile · #26 of 38 open · #57 of 76 all
#26 of 38
#57 of 76
Lab-reported
Epoch AI · Feb 2024
TriviaQA
TriviaQA is a large set of trivia questions paired with evidence documents, testing a model's factual recall and reading comprehension across a wide range of everyday topics.
53.2%
53.2%
median 73.7%open leader 87.6%
6th percentile · #18 of 19 open · #38 of 39 all
#18 of 19
#38 of 39
Lab-reported
Epoch AI · Feb 2024

Scores aggregated from public benchmark sources (Epoch AI: 9 boards, TIGER-Lab: 1 board) — each linked from the benchmark pages. llmrun does not run these benchmarks. Percentiles and ranks are among open models in the llmrun catalogue; the median and leader marks refer to the same field.