Model comparison
Gemma 3 4B vs Qwen2.5 72B Instruct
Qwen2.5 72B Instruct is the stronger model overall, scoring 31.9 to 28.1 on the Noometry Index. Gemma 3 4B costs 49× less per token, which makes it the better buy when Qwen2.5 72B Instruct's lead doesn't matter for your workload.
Last verified . 17 shared benchmarks.
Google
Alibaba (Qwen)Summary
- They share 17 benchmarks with published results for both. Gemma 3 4B scores higher in 2 categories and Qwen2.5 72B Instruct in 7 categories; 7 gaps are clear of the uncertainty.
- The widest gap is in knowledge, where Qwen2.5 72B Instruct leads 27.0 to 11.8.
- The biggest single-benchmark swing is GPQA Diamond: 23.2% for Gemma 3 4B and 49.1% for Qwen2.5 72B Instruct.
- Gemma 3 4B is cheaper at $0.04 / $0.08 per million input/output tokens, against $1.40 / $5.60 for Qwen2.5 72B Instruct.
Side by side
| Gemma 3 4B | Qwen2.5 72B Instruct | |
|---|---|---|
| Provider | Alibaba (Qwen) | |
| Noometry Index | 28.1 | 31.9 |
| Released | 2025-03-12 | 2024-09 |
| Weights | Open | Open |
| Context window | 131K | 131K |
| Max output | 4K | 8K |
| Input $ / M tokens | $0.04 | $1.40 |
| Output $ / M tokens | $0.08 | $5.60 |
| Results tracked | 22 | 43 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding Gemma 3 4B leads
Gemma 3 4B: 35.9 (#215), Qwen2.5 72B Instruct: 33.2 (#260)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Coding | 1230 | 1292 |
| WeirdML | — | 16% |
| BigCodeBench Instruct | — | 45.8% |
| BigCodeBench Complete | — | 55.9% |
Agentic & Tool Use Qwen2.5 72B Instruct leads
Gemma 3 4B: 20.9 (#142), Qwen2.5 72B Instruct: 22.1 (#133)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| Berkeley Function Calling Leaderboard | 19.6% | — |
| TheAgentCompany | — | 5.7% |
| BALROG | — | 16.2% |
| METR Time Horizons | — | 35.8% |
Reasoning Qwen2.5 72B Instruct leads
Gemma 3 4B: 13.2 (#335), Qwen2.5 72B Instruct: 22.3 (#199)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Hard Prompts | 1253 | 1271 |
| DTBench | 50.9% | 62.9% |
| LMCA | 2.8% | 13.4% |
| Epoch Capabilities Index | 116.02 | 129 |
| Kagi LLM Benchmark | 25.2% | — |
| Chess Puzzles | 0% | — |
| BIG-Bench Hard | — | 79.8% |
| ForecastBench | — | 57.5 |
| HellaSwag | — | 84.8% |
| PIQA | — | 82.6% |
| WinoGrande | — | 82.3% |
Math Qwen2.5 72B Instruct leads
Gemma 3 4B: 16.8 (#292), Qwen2.5 72B Instruct: 19.3 (#287)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 7.5% | 8.1% |
| LMArena Math | 1239 | 1283 |
| Omni-MATH | — | 33% |
| MATH Level 5 | — | 63.2% |
Knowledge Qwen2.5 72B Instruct leads
Gemma 3 4B: 11.8 (#299), Qwen2.5 72B Instruct: 27.0 (#253)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| GPQA Diamond | 23.2% | 49.1% |
| LMArena Expert | 1223 | 1245 |
| MMLU-Pro | — | 63.1% |
| Confabulations | — | 19.1% |
| Vectara Hallucination Rate | 6.4% | — |
| GPQA (HELM) | — | 42.6% |
| ARC (AI2) Challenge | — | 94.5% |
| MMLU | — | 85.3% |
| TriviaQA | — | 71.9% |
Multilingual Gemma 3 4B leads
Gemma 3 4B: 42.5 (#194), Qwen2.5 72B Instruct: 41.0 (#213)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Non-English | 1273 | 1252 |
| LMArena German | 1281 | 1234 |
| LMArena Russian | 1294 | 1264 |
| LMArena Chinese | — | 1272 |
| LMArena French | — | 1280 |
| LMArena Japanese | — | 1180 |
| LMArena Korean | — | 1188 |
| LMArena Spanish | — | 1256 |
Instruction Following Too close to call
Gemma 3 4B: 65.2 (#225), Qwen2.5 72B Instruct: 65.5 (#221)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Instruction Following | 1239 | 1254 |
| IFEval | — | 80.6% |
Long Context Too close to call
Gemma 3 4B: 38.7 (#194), Qwen2.5 72B Instruct: 38.9 (#188)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Longer Query | 1273 | 1282 |
Writing & Preference Qwen2.5 72B Instruct leads
Gemma 3 4B: 42.0 (#239), Qwen2.5 72B Instruct: 46.7 (#215)
| Benchmark | Gemma 3 4B | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Text | 1291 | 1269 |
| LMArena Creative Writing | 1271 | 1221 |
| LMArena Multi-Turn | 1255 | 1272 |
| EQ-Bench Creative Writing | 1068 | — |
| WildBench | — | 80.2% |
Frequently asked questions
Is Gemma 3 4B better than Qwen2.5 72B Instruct?
Qwen2.5 72B Instruct is the stronger model overall, scoring 31.9 to 28.1 on the Noometry Index. Gemma 3 4B costs 49× less per token, which makes it the better buy when Qwen2.5 72B Instruct's lead doesn't matter for your workload.
Which is cheaper, Gemma 3 4B or Qwen2.5 72B Instruct?
Gemma 3 4B is cheaper. It lists at $0.04 per million input tokens and $0.08 per million output tokens; Qwen2.5 72B Instruct lists at $1.40 and $5.60.
Is Gemma 3 4B or Qwen2.5 72B Instruct better for coding?
Gemma 3 4B scores higher on coding benchmarks: 35.9 versus 33.2 in the Noometry coding category.
Which has the bigger context window?
Both accept 131K tokens.
How many benchmarks do Gemma 3 4B and Qwen2.5 72B Instruct share?
17 benchmarks have published results for both models. Gemma 3 4B has 22 scored results on Noometry and Qwen2.5 72B Instruct has 43.