Model comparison
DeepSeek-R1 vs Step 3
DeepSeek-R1 is the stronger model overall, scoring 42.0 to 39.9 on the Noometry Index.
Last verified . 16 shared benchmarks.
DeepSeek
StepFunSummary
- They share 16 benchmarks with published results for both. DeepSeek-R1 scores higher in 7 categories and Step 3 in 1 category; 8 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where Step 3 leads 27.6 to 18.7.
- The biggest single-benchmark swing is Kagi LLM Benchmark: 69.4% for DeepSeek-R1 and 62.3% for Step 3.
- Step 3 has downloadable open weights; the other is API-only.
Side by side
| DeepSeek-R1 | Step 3 | |
|---|---|---|
| Provider | DeepSeek | StepFun |
| Noometry Index | 42.0 | 39.9 |
| Released | 2025-01-20 | — |
| Weights | Proprietary | Open |
| Context window | 164K | — |
| Max output | 64K | — |
| Input $ / M tokens | $0.50 | — |
| Output $ / M tokens | $2.15 | — |
| Results tracked | 52 | 17 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek-R1 leads
DeepSeek-R1: 45.9 (#63), Step 3: 39.5 (#144)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Coding | 1427 | 1367 |
| Aider Polyglot | 71.4% | — |
| SciCode | 35.7% | — |
| WeirdML | 41.6% | — |
| LiveBench Coding | 66.7% | — |
| ALE-Bench | 804.12 | — |
| AlgoTune | 1.7 | — |
Agentic & Tool Use Not comparable
DeepSeek-R1: 30.4 (#68), Step 3: —
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| DeepResearch Bench | 35.1% | — |
| BALROG | 34.9% | — |
| METR Time Horizons | 53.8% | — |
Reasoning Step 3 leads
DeepSeek-R1: 18.7 (#253), Step 3: 27.6 (#102)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| Kagi LLM Benchmark | 69.4% | 62.3% |
| LMArena Hard Prompts | 1415 | 1355 |
| ARC-AGI-2 | 1.3% | — |
| SimpleBench | 40.8% | — |
| ARC-AGI-1 | 21.2% | — |
| CritPt | 1.1% | — |
| LiveBench Reasoning | 83.2% | — |
| LiveBench Data Analysis | 69.8% | — |
| Epoch Capabilities Index | 141.3 | — |
| ForecastBench | 60 | — |
| LiveBench | 71.6% | — |
Math DeepSeek-R1 leads
DeepSeek-R1: 42.0 (#70), Step 3: 34.9 (#148)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Math | 1401 | 1366 |
| OTIS Mock AIME 2024-2025 | 66.4% | — |
| Omni-MATH | 42.4% | — |
| LiveBench Math | 80.7% | — |
| MATH Level 5 | 96.6% | — |
Knowledge DeepSeek-R1 leads
DeepSeek-R1: 44.3 (#81), Step 3: 36.2 (#161)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Expert | 1394 | 1333 |
| GPQA Diamond | 76.3% | — |
| MMLU-Pro | 79.3% | — |
| Confabulations | 12.7% | — |
| Vectara Hallucination Rate | 11.3% | — |
| GPQA (HELM) | 66.6% | — |
Multimodal Not comparable
DeepSeek-R1: —, Step 3: 34.9 (#81)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Vision | — | 1176 |
Multilingual DeepSeek-R1 leads
DeepSeek-R1: 52.6 (#80), Step 3: 46.5 (#154)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Non-English | 1412 | 1327 |
| LMArena Chinese | 1443 | 1396 |
| LMArena German | 1405 | 1371 |
| LMArena Korean | 1359 | 1271 |
| LMArena Russian | 1423 | 1331 |
| LMArena Spanish | 1410 | 1372 |
| LMArena French | 1415 | — |
| LMArena Japanese | 1394 | — |
Instruction Following DeepSeek-R1 leads
DeepSeek-R1: 72.2 (#138), Step 3: 70.6 (#159)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Instruction Following | 1382 | 1332 |
| LiveBench Instruction Following | 80.5% | — |
| IFEval | 78.4% | — |
Long Context DeepSeek-R1 leads
DeepSeek-R1: 45.4 (#36), Step 3: 40.3 (#152)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Longer Query | 1391 | 1326 |
| Fiction.LiveBench | 75% | — |
Writing & Preference DeepSeek-R1 leads
DeepSeek-R1: 61.5 (#84), Step 3: 54.5 (#146)
| Benchmark | DeepSeek-R1 | Step 3 |
|---|---|---|
| LMArena Text | 1428 | 1351 |
| LMArena Creative Writing | 1406 | 1322 |
| LMArena Multi-Turn | 1404 | 1340 |
| Short-Story Creative Writing | 83% | — |
| EQ-Bench Creative Writing | 1500 | — |
| WildBench | 82.8% | — |
| LiveBench Language | 48.5% | — |
Frequently asked questions
Is DeepSeek-R1 better than Step 3?
DeepSeek-R1 is the stronger model overall, scoring 42.0 to 39.9 on the Noometry Index.
Is DeepSeek-R1 or Step 3 better for coding?
DeepSeek-R1 scores higher on coding benchmarks: 45.9 versus 39.5 in the Noometry coding category.
How many benchmarks do DeepSeek-R1 and Step 3 share?
16 benchmarks have published results for both models. DeepSeek-R1 has 52 scored results on Noometry and Step 3 has 17.