Model comparison
Gemma 3 12B vs Gemma 4 31B IT
Gemma 4 31B IT is the stronger model overall, scoring 43.5 to 32.1 on the Noometry Index. Gemma 3 12B costs 2.0× less per token, which makes it the better buy when Gemma 4 31B IT's lead doesn't matter for your workload.
Last verified . 21 shared benchmarks.
GoogleSummary
- They share 21 benchmarks with published results for both. Gemma 3 12B scores higher in 0 categories and Gemma 4 31B IT in 8 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where Gemma 4 31B IT leads 43.2 to 22.3.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 16.7% for Gemma 3 12B and 73.3% for Gemma 4 31B IT.
- Gemma 3 12B is cheaper at $0.05 / $0.15 per million input/output tokens, against $0.09 / $0.34 for Gemma 4 31B IT.
- Gemma 4 31B IT accepts more context: 262K tokens versus 131K.
Side by side
| Gemma 3 12B | Gemma 4 31B IT | |
|---|---|---|
| Provider | ||
| Noometry Index | 32.1 | 43.5 |
| Released | 2025-03-12 | 2026-04-02 |
| Weights | Open | Open |
| Context window | 131K | 262K |
| Max output | 8K | 33K |
| Input $ / M tokens | $0.05 | $0.09 |
| Output $ / M tokens | $0.15 | $0.34 |
| Results tracked | 24 | 35 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding Gemma 4 31B IT leads
Gemma 3 12B: 31.7 (#280), Gemma 4 31B IT: 42.3 (#108)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| SciCode | 17.4% | 43.4% |
| LMArena Coding | 1281 | 1459 |
| LMArena WebDev | — | 1366 |
| WeirdML | — | 52.3% |
| ALE-Bench | — | 925.5 |
Agentic & Tool Use Not comparable
Gemma 3 12B: 25.5 (#108), Gemma 4 31B IT: —
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| Berkeley Function Calling Leaderboard | 30.4% | — |
Reasoning Gemma 4 31B IT leads
Gemma 3 12B: 15.7 (#313), Gemma 4 31B IT: 27.2 (#122)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| CritPt | 0% | 1.4% |
| Chess Puzzles | 0% | 5% |
| LMArena Hard Prompts | 1309 | 1448 |
| DTBench | 48.8% | 82.7% |
| LMCA | 4.5% | 39.3% |
| Epoch Capabilities Index | 123.5 | 142.74 |
| Kagi LLM Benchmark | — | 63.5% |
| NYT Connections (extended) | — | 70.6% |
| Thematic Generalization | — | 53% |
| Surface Evolver Bench | — | 30.6% |
Math Gemma 4 31B IT leads
Gemma 3 12B: 22.3 (#279), Gemma 4 31B IT: 43.2 (#81)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 16.7% | 73.3% |
| LMArena Math | 1307 | 1465 |
Knowledge Gemma 4 31B IT leads
Gemma 3 12B: 26.5 (#257), Gemma 4 31B IT: 37.9 (#151)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| GPQA Diamond | 39.5% | 75.8% |
| Vectara Hallucination Rate | 4.4% | 7.4% |
| LMArena Expert | 1248 | 1465 |
| SimpleQA Verified | — | 10.4% |
Multimodal Not comparable
Gemma 3 12B: —, Gemma 4 31B IT: 41.6 (#34)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| LMArena Vision | — | 1277 |
| LMArena Document | — | 1425 |
| MindCube | 46.7% | — |
Multilingual Gemma 4 31B IT leads
Gemma 3 12B: 45.7 (#165), Gemma 4 31B IT: 53.8 (#57)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| LMArena Non-English | 1318 | 1431 |
| LMArena Russian | 1335 | 1460 |
| LMArena Chinese | — | 1476 |
| LMArena French | — | 1435 |
| LMArena German | 1370 | — |
| LMArena Spanish | — | 1444 |
Instruction Following Gemma 4 31B IT leads
Gemma 3 12B: 68.6 (#186), Gemma 4 31B IT: 75.5 (#61)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| LMArena Instruction Following | 1299 | 1433 |
Long Context Gemma 4 31B IT leads
Gemma 3 12B: 40.0 (#162), Gemma 4 31B IT: 44.2 (#71)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| LMArena Longer Query | 1317 | 1446 |
Writing & Preference Gemma 4 31B IT leads
Gemma 3 12B: 47.5 (#209), Gemma 4 31B IT: 60.5 (#96)
| Benchmark | Gemma 3 12B | Gemma 4 31B IT |
|---|---|---|
| LMArena Text | 1334 | 1443 |
| LMArena Creative Writing | 1331 | 1415 |
| EQ-Bench Creative Writing | 1126 | 1368 |
| LMArena Multi-Turn | 1334 | 1452 |
| EQ-Bench 4 | — | 1120 |
Frequently asked questions
Is Gemma 3 12B better than Gemma 4 31B IT?
Gemma 4 31B IT is the stronger model overall, scoring 43.5 to 32.1 on the Noometry Index. Gemma 3 12B costs 2.0× less per token, which makes it the better buy when Gemma 4 31B IT's lead doesn't matter for your workload.
Which is cheaper, Gemma 3 12B or Gemma 4 31B IT?
Gemma 3 12B is cheaper. It lists at $0.05 per million input tokens and $0.15 per million output tokens; Gemma 4 31B IT lists at $0.09 and $0.34.
Is Gemma 3 12B or Gemma 4 31B IT better for coding?
Gemma 4 31B IT scores higher on coding benchmarks: 42.3 versus 31.7 in the Noometry coding category.
Which has the bigger context window?
Gemma 4 31B IT does, with 262K tokens against 131K.
How many benchmarks do Gemma 3 12B and Gemma 4 31B IT share?
21 benchmarks have published results for both models. Gemma 3 12B has 24 scored results on Noometry and Gemma 4 31B IT has 35.