Gemma 3 27B vs Gemma 4 12B
Wins 0 of 8 areas
—
Wins 8 of 8 areas
Coding · Agents · Reasoning · Facts · Images and charts · Math · Long documents · Following instructions
Gemma 4 12B is the stronger all-rounder.Gemma 3 27B is cheaper.
Scores updated · 27 tests both models report · How we compare
Where each one wins
Tests won in each of the eight areas we test. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- CodingWriting and fixing software04Gemma 4 12B4 of 4 tests
- ReasoningHard problems that need careful thinking02Gemma 4 12B2 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up02Gemma 4 12B2 of 2 tests
- Long documentsFinding answers in very long texts02Gemma 4 12B2 of 2 tests
- AgentsCarrying out multi-step tasks on its own01Gemma 4 12B1 of 2 tests · 1 tie
- Images and chartsUnderstanding pictures, charts and video01Gemma 4 12B1 of 1 test
- MathCompetition and research-level math01Gemma 4 12B1 of 1 test
- Following instructionsDoing exactly what it is asked01Gemma 4 12B1 of 1 test
Images and charts, math and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
Gemma 3 27B costs 40% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where Gemma 4 12B pulls ahead
- US invitational high-school math exam problemsAIME 2026+56.7points ahead
- Reasons across sets of long documentsAA-LCR+56.4points ahead
- Recent programming contest problemsLiveCodeBench v6+42.9points ahead
Where Gemma 3 27B pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 27 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingGemma 4 12B
- LiveCodeBench v6Gemma 4 12B by 42.929.172+42.9
- Terminal-Bench 2.1Gemma 4 12B by 22.84.527.3+22.8
- SciCodeGemma 4 12B by 14.923.338.2+14.9
- Terminal-Bench HardGemma 4 12B by 14.43.818.2+14.4
ReasoningGemma 4 12B
- GPQA DiamondGemma 4 12B by 32.542.875.3+32.5
- Humanity's Last ExamGemma 4 12B by 11.34.415.7+11.3
- CritPttie00tie
FactsGemma 4 12B
- AA-Omniscience · Non-hallucinationGemma 4 12B by 11.17.919+11.1
- AA-Omniscience · AccuracyGemma 4 12B by 2.612.915.6+2.6
Long documentsGemma 4 12B
- AA-LCRGemma 4 12B by 56.47.363.7+56.4
- MRCR v2 8 needle 128k (average)Gemma 4 12B by 29.913.543.4+29.9
Following instructionsGemma 4 12B
- IFBenchGemma 4 12B by 41.731.873.5+41.7
Other results11 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- Codeforces ELOGemma 4 12B by 1549 rating points1101659+1549 rating
- τ²-BenchGemma 4 12B by 52.816.269+52.8
- BIG-Bench Extra HardGemma 4 12B by 33.719.353+33.7
- MATH-VisionGemma 4 12B by 33.74679.7+33.7
- τ²-Bench Telecom (AA run)Gemma 4 12B by 25.810.536.3+25.8
- Artificial Analysis Coding IndexGemma 4 12B by 20.910.131+20.9
- AA-OmniscienceGemma 4 12B by 14.5-67.2-52.7+14.5
- MMMLUGemma 4 12B by 12.770.783.4+12.7
- MMLU-ProGemma 4 12B by 9.667.677.2+9.6
- AA IntelligenceGemma 4 12B by 9.34.914.2+9.3
- OmniDocBench 1.5 (average edit distance, lower is better)lower is bettertie0.40.2tie
Questions people ask
Which is better, Gemma 3 27B or Gemma 4 12B?
Gemma 4 12B wins all eight areas we test: coding, agents, reasoning, facts, images and charts, math, long documents and following instructions. Gemma 3 27B wins none, but costs 40% less.
Which is better for coding?
Gemma 4 12B. It wins 4 of the 4 coding tests both models report; Gemma 3 27B wins none.
Which is cheaper?
Gemma 3 27B costs $0.08 per million input tokens and $0.16 per million output tokens; Gemma 4 12B costs $0.10 and $0.30. That makes Gemma 3 27B about 40% cheaper for the same work.
How do you compare the two?
We use the 27 benchmark tests both models have published scores on. The verdict counts the 16 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 11 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.