# DeepSeek V4 Flash vs Qwen2.5 72B Instruct

> DeepSeek V4 Flash is the stronger model overall, scoring 53.6 to 31.5 on the Noometry Index.

- Canonical page: https://noometry.com/compare/deepseek-v4-flash-vs-qwen2-5-72b-instruct
- Last updated: 2026-10-10
- Shared benchmarks: 23

## Summary

- They share 23 benchmarks with published results for both. DeepSeek V4 Flash scores higher in 8 categories and Qwen2.5 72B Instruct in 0 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where DeepSeek V4 Flash leads 60.0 to 18.1.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 94.4% for DeepSeek V4 Flash and 8.1% for Qwen2.5 72B Instruct.
- DeepSeek V4 Flash is cheaper at $0.15 / $0.60 per million input/output tokens, against $1.40 / $5.60 for Qwen2.5 72B Instruct.
- DeepSeek V4 Flash accepts more context: 1M tokens versus 131K.

## Snapshot

| | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| Provider | DeepSeek | Alibaba (Qwen) |
| Noometry Index | 53.6 | 31.5 |
| Rank | 32 | 259 |
| Context | 1M | 131K |
| Input $/M | $0.15 | $1.40 |
| Output $/M | $0.60 | $5.60 |
| Weights | Open | Open |

## Coding

- DeepSeek V4 Flash: 47.6 (#56)
- Qwen2.5 72B Instruct: 32.8 (#256)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| WeirdML | 63% | 16% |
| LMArena Coding | 1457 | 1293 |
| FrontierCode | 18.8% | — |
| LMArena WebDev | 1581 | — |
| SciCode | 49.9% | — |
| BigCodeBench Instruct | — | 45.8% |
| BigCodeBench Complete | — | 55.9% |
| ALE-Bench | 1,306 | — |

## Agentic & Tool Use

- DeepSeek V4 Flash: —
- Qwen2.5 72B Instruct: 22.0 (#124)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| TheAgentCompany | — | 5.7% |
| BALROG | — | 16.2% |
| METR Time Horizons | — | 35.8% |

## Reasoning

- DeepSeek V4 Flash: 53.7 (#25)
- Qwen2.5 72B Instruct: 21.4 (#197)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Hard Prompts | 1443 | 1271 |
| DTBench | 90.9% | 62.9% |
| LMCA | 41.7% | 13.4% |
| Epoch Capabilities Index | 154.51 | 129.01 |
| ARC-AGI-2 | 61.4% | — |
| SimpleBench | 61.1% | — |
| Kagi LLM Benchmark | 52.2% | — |
| NYT Connections (extended) | 89.6% | — |
| ARC-AGI-1 | 89% | — |
| CritPt | 16.6% | — |
| Chess Puzzles | 33% | — |
| Mystery Game Puzzles | 34% | — |
| BIG-Bench Hard | — | 79.8% |
| ForecastBench | — | 57.5 |
| HellaSwag | — | 84.8% |
| PIQA | — | 82.6% |
| WinoGrande | — | 82.3% |

## Math

- DeepSeek V4 Flash: 60.0 (#31)
- Qwen2.5 72B Instruct: 18.1 (#281)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 94.4% | 8.1% |
| LMArena Math | 1427 | 1283 |
| FrontierMath (Tiers 1-3) | 57.5% | — |
| FrontierMath Tier 4 | 24.4% | — |
| MathArena Final-Answer Competitions | 76.5% | — |
| ProofBench | 56% | — |
| Omni-MATH | — | 33% |
| MATH Level 5 | — | 63.2% |

## Knowledge

- DeepSeek V4 Flash: 55.5 (#43)
- Qwen2.5 72B Instruct: 26.9 (#247)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| GPQA Diamond | 91% | 49.1% |
| LMArena Expert | 1441 | 1245 |
| SimpleQA Verified | 33.6% | — |
| MMLU-Pro | — | 63.1% |
| Confabulations | — | 19.1% |
| GPQA (HELM) | — | 42.6% |
| ARC (AI2) Challenge | — | 94.5% |
| MMLU | — | 85.3% |
| TriviaQA | — | 71.9% |

## Multilingual

- DeepSeek V4 Flash: 53.2 (#71)
- Qwen2.5 72B Instruct: 41.2 (#208)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Non-English | 1420 | 1252 |
| LMArena Chinese | 1467 | 1272 |
| LMArena French | 1438 | 1279 |
| LMArena German | 1418 | 1234 |
| LMArena Japanese | 1403 | 1180 |
| LMArena Korean | 1383 | 1188 |
| LMArena Russian | 1428 | 1264 |
| LMArena Spanish | 1437 | 1255 |

## Instruction Following

- DeepSeek V4 Flash: 75.1 (#76)
- Qwen2.5 72B Instruct: 65.7 (#216)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Instruction Following | 1421 | 1254 |
| IFEval | — | 80.6% |

## Long Context

- DeepSeek V4 Flash: 43.8 (#81)
- Qwen2.5 72B Instruct: 38.9 (#183)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Longer Query | 1435 | 1282 |

## Writing & Preference

- DeepSeek V4 Flash: 64.0 (#59)
- Qwen2.5 72B Instruct: 46.9 (#210)

| Benchmark | DeepSeek V4 Flash | Qwen2.5 72B Instruct |
|---|---|---|
| LMArena Text | 1432 | 1269 |
| LMArena Creative Writing | 1403 | 1221 |
| LMArena Multi-Turn | 1449 | 1272 |
| EQ-Bench Creative Writing | 1559 | — |
| WildBench | — | 80.2% |

## FAQ

### Is DeepSeek V4 Flash better than Qwen2.5 72B Instruct?

DeepSeek V4 Flash is the stronger model overall, scoring 53.6 to 31.5 on the Noometry Index.

### Which is cheaper, DeepSeek V4 Flash or Qwen2.5 72B Instruct?

DeepSeek V4 Flash is cheaper. It lists at $0.15 per million input tokens and $0.60 per million output tokens; Qwen2.5 72B Instruct lists at $1.40 and $5.60.

### Is DeepSeek V4 Flash or Qwen2.5 72B Instruct better for coding?

DeepSeek V4 Flash scores higher on coding benchmarks: 47.6 versus 32.8 in the Noometry coding category.

### Which has the bigger context window?

DeepSeek V4 Flash does, with 1M tokens against 131K.

### How many benchmarks do DeepSeek V4 Flash and Qwen2.5 72B Instruct share?

23 benchmarks have published results for both models. DeepSeek V4 Flash has 41 scored results on Noometry and Qwen2.5 72B Instruct has 43.