Google: Gemma 3 4B
Google: Gemma 3 4B is google's model. On the un-leaked meo benchmark it ranks #72 of 82 with 47.3% accuracy and an Effective Value (π) of 0.0001 β earning a correct answer for about $0.0001 in 36s each.
as of 2026-10-18
Per-domain accuracy
116 correct of 250 questions asked across 11 domains Β· answer rate 100% Β· macro accuracy 38.1%. Headline accuracy is micro-averaged per meoβs methodology; per-domain question counts (correct/asked) are shown so you can judge significance.
Efficiency
- Cost per correct
- $0.0001
- Seconds per correct
- 36
- Tokens per correct
- 1,081
- Correct answers
- 116 / 250
- Total benchmark cost
- $0.02
- Avg tokens / item
- 502
Aggregated third-party benchmarks
Artificial Analysis (artificialanalysis.ai). Redistribution requires an AA commercial license.
Model details
- Lab
- Released
- 2025-03-13
- Weights
- Open
- Context
- 131K
- Max output
- 16K
- Knowledge cutoff
- 2024-08-31
- Price in
- $0.05/M
- Price out
- $0.10/M
- Modality
- text+image->text
Sources: openrouter Β· lmarena Β· artificial_analysis Β· meo
FAQ
How accurate is Google: Gemma 3 4B?
Google: Gemma 3 4B scores 47.3% overall on the meo benchmark (rank #72 of 82), an un-leaked private holdout graded objective-first.
What is Google: Gemma 3 4B's Effective Value?
Its Effective Value (π) is 0.0001 at N=10, Ο=1, Ξ΄=1.5 β a fused metric of accuracy, speed, cost, and error-cascade. π rankings shift with task depth.
How much does Google: Gemma 3 4B cost per correct answer?
About $0.0001 per correct answer (1,081 tokens, 36 seconds each) at benchmark-time pricing via OpenRouter.
Is Google: Gemma 3 4B open weights?
Yes β Google: Gemma 3 4B is released with open weights.
Benchmark updates affecting this model
- 2026-06-09majorlogic_math_cs ground truth was broken β culled and rebuilt with solver-verified items
A 5-lab frontier jury revealed that 29 of 37 LLM-authored logic_math_cs items had broken ground truth: on most, five independent frontier models unanimously concluded the clue sets were contradictory while the stored answer claimed a number, and on eight, all five agreed on the SAME different number (author arithmetic errors). Models were being marked wrong for being right. The 29 were retired (full audit published in the repo) and the domain rebuilt with generator-as-oracle items: an embedded constraint solver PROVES each puzzle has exactly one solution (or provably none), verified by an independent exhaustive enumerator. A 3-round difficulty study found frontier models saturate consistent human-scale logic puzzles, so items use anonymous global constraints and carry a measured difficulty gradient.
- 2026-06-08majorCorrected base_bias grading (LLM-equivalence judge fix)
The LLM-equivalence grader uses a reasoning model whose verdict token budget was too small, so its YES/NO answer was emitted as empty and silently counted as a failure. This understated the base_bias domain to ~7% for every model when the true accuracy is ~85-90%. We raised the judge budget, excluded its reasoning tokens from the verdict, switched to verbosity-robust conclusion matching, and re-graded the whole board.
Embed this ranking
Show Google: Gemma 3 4Bβs meo rank on your site β the badge links back here.
Related: all Google models Β· Base-rate bias domain Β· best Effective Value Β· biggest movers Β· methodology Β· full leaderboard.