Model comparison
Phi-4 vs Qwen2.5-Coder-32B
Qwen2.5-Coder-32B is the stronger model overall, scoring 33.4 to 31.2 on the Noometry Index. Phi-4 costs 8.5× less per token, which makes it the better buy when Qwen2.5-Coder-32B's lead doesn't matter for your workload.
Last verified . 23 shared benchmarks.
Microsoft
Alibaba (Qwen)Summary
- They share 23 benchmarks with published results for both. Phi-4 scores higher in 1 category and Qwen2.5-Coder-32B in 7 categories; 5 gaps are clear of the uncertainty.
- The widest gap is in math, where Qwen2.5-Coder-32B leads 33.3 to 20.8.
- The biggest single-benchmark swing is LiveBench Coding: 30.7% for Phi-4 and 56.9% for Qwen2.5-Coder-32B.
- Phi-4 is cheaper at $0.07 / $0.14 per million input/output tokens, against $0.66 / $1 for Qwen2.5-Coder-32B.
- Phi-4 accepts more context: 128K tokens versus 33K.
Side by side
| Phi-4 | Qwen2.5-Coder-32B | |
|---|---|---|
| Provider | Microsoft | Alibaba (Qwen) |
| Noometry Index | 31.2 | 33.4 |
| Released | 2024-12-11 | 2024-09-18 |
| Weights | Open | Open |
| Context window | 128K | 33K |
| Max output | 4K | 29K |
| Input $ / M tokens | $0.07 | $0.66 |
| Output $ / M tokens | $0.14 | $1 |
| Results tracked | 37 | 31 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding Phi-4 leads
Phi-4: 34.4 (#239), Qwen2.5-Coder-32B: 22.6 (#333)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| BigCodeBench Instruct | 45.5% | 49% |
| LiveBench Coding | 30.7% | 56.9% |
| LMArena Coding | 1231 | 1276 |
| BigCodeBench Complete | 55.4% | 58% |
| SWE-bench Verified (bash only) | — | 9% |
| Aider Polyglot | — | 16.4% |
| HumanEval+ | — | 87.2% |
| MBPP+ | — | 77% |
Agentic & Tool Use Not comparable
Phi-4: 22.8 (#128), Qwen2.5-Coder-32B: —
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| Berkeley Function Calling Leaderboard | 28.8% | — |
| BALROG | 11.6% | — |
Reasoning Qwen2.5-Coder-32B leads
Phi-4: 17.7 (#291), Qwen2.5-Coder-32B: 21.2 (#225)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| LiveBench Reasoning | 47.8% | 42.1% |
| LMArena Hard Prompts | 1220 | 1251 |
| LiveBench Data Analysis | 45.2% | 49.9% |
| Epoch Capabilities Index | 130.42 | 119.49 |
| LiveBench | 41.6% | 46.2% |
| Chess Puzzles | 1% | — |
| HellaSwag | — | 83% |
| WinoGrande | — | 80.8% |
Math Qwen2.5-Coder-32B leads
Phi-4: 20.8 (#285), Qwen2.5-Coder-32B: 33.3 (#204)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| LiveBench Math | 42% | 46.6% |
| LMArena Math | 1246 | 1251 |
| OTIS Mock AIME 2024-2025 | 13.8% | — |
| MATH Level 5 | 64.9% | — |
| GSM8K | — | 93% |
Knowledge Too close to call
Phi-4: 32.6 (#209), Qwen2.5-Coder-32B: 33.4 (#203)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| LMArena Expert | 1203 | 1221 |
| MMLU | 84.8% | 79.1% |
| GPQA Diamond | 56.1% | — |
| Confabulations | 29.4% | — |
| Vectara Hallucination Rate | 3.7% | — |
| ARC (AI2) Challenge | — | 70.5% |
Multilingual Too close to call
Phi-4: 37.2 (#237), Qwen2.5-Coder-32B: 37.8 (#235)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| LMArena Non-English | 1197 | 1205 |
| LMArena Chinese | 1212 | 1222 |
| LMArena Russian | 1209 | 1228 |
| LMArena French | 1224 | — |
| LMArena German | 1222 | — |
| LMArena Japanese | 1158 | — |
| LMArena Korean | 1151 | — |
| LMArena Spanish | 1234 | — |
Instruction Following Too close to call
Phi-4: 60.4 (#251), Qwen2.5-Coder-32B: 61.4 (#245)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| LiveBench Instruction Following | 58.4% | 58.7% |
| LMArena Instruction Following | 1201 | 1223 |
Long Context Qwen2.5-Coder-32B leads
Phi-4: 36.9 (#226), Qwen2.5-Coder-32B: 38.0 (#208)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| LMArena Longer Query | 1217 | 1251 |
Writing & Preference Qwen2.5-Coder-32B leads
Phi-4: 40.5 (#244), Qwen2.5-Coder-32B: 41.6 (#240)
| Benchmark | Phi-4 | Qwen2.5-Coder-32B |
|---|---|---|
| LMArena Text | 1217 | 1230 |
| LMArena Creative Writing | 1182 | 1174 |
| LMArena Multi-Turn | 1206 | 1222 |
| LiveBench Language | 25.6% | 23.3% |
| Short-Story Creative Writing | 62.6% | — |
Frequently asked questions
Is Phi-4 better than Qwen2.5-Coder-32B?
Qwen2.5-Coder-32B is the stronger model overall, scoring 33.4 to 31.2 on the Noometry Index. Phi-4 costs 8.5× less per token, which makes it the better buy when Qwen2.5-Coder-32B's lead doesn't matter for your workload.
Which is cheaper, Phi-4 or Qwen2.5-Coder-32B?
Phi-4 is cheaper. It lists at $0.07 per million input tokens and $0.14 per million output tokens; Qwen2.5-Coder-32B lists at $0.66 and $1.
Is Phi-4 or Qwen2.5-Coder-32B better for coding?
Phi-4 scores higher on coding benchmarks: 34.4 versus 22.6 in the Noometry coding category.
Which has the bigger context window?
Phi-4 does, with 128K tokens against 33K.
How many benchmarks do Phi-4 and Qwen2.5-Coder-32B share?
23 benchmarks have published results for both models. Phi-4 has 37 scored results on Noometry and Qwen2.5-Coder-32B has 31.