Model comparison
DeepSeek V4 Pro vs Tulu 3 (Tülu 3) 70B
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 33.0 on the Noometry Index.
Last verified . 13 shared benchmarks.
DeepSeek
Allen Institute for AI (Ai2)Summary
- They share 13 benchmarks with published results for both. DeepSeek V4 Pro scores higher in 8 categories and Tulu 3 (Tülu 3) 70B in 0 categories; 8 gaps are clear of the uncertainty.
- The widest gap is in math, where DeepSeek V4 Pro leads 64.8 to 14.2.
- The biggest single-benchmark swing is OTIS Mock AIME 2024-2025: 98.6% for DeepSeek V4 Pro and 4.4% for Tulu 3 (Tülu 3) 70B.
Side by side
| DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B | |
|---|---|---|
| Provider | DeepSeek | Allen Institute for AI (Ai2) |
| Noometry Index | 54.3 | 33.0 |
| Released | 2026-04-24 | 2024-11-21 |
| Weights | Open | Open |
| Context window | 1M | — |
| Max output | 393K | — |
| Input $ / M tokens | $0.66 | — |
| Output $ / M tokens | $1.98 | — |
| Results tracked | 48 | 14 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding DeepSeek V4 Pro leads
DeepSeek V4 Pro: 52.4 (#34), Tulu 3 (Tülu 3) 70B: 36.0 (#214)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| LMArena Coding | 1470 | 1235 |
| SWE-bench Verified | 77.6% | — |
| FrontierCode | 28.6% | — |
| LMArena WebDev | 1582 | — |
| SciCode | 51% | — |
| WeirdML | 66.2% | — |
| ALE-Bench | 1,403 | — |
Agentic & Tool Use Not comparable
DeepSeek V4 Pro: 32.8 (#58), Tulu 3 (Tülu 3) 70B: —
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| APEX-Agents | 47.3% | — |
| Vending-Bench 2 | 3,285 | — |
Reasoning DeepSeek V4 Pro leads
DeepSeek V4 Pro: 56.5 (#24), Tulu 3 (Tülu 3) 70B: 23.9 (#169)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| LMArena Hard Prompts | 1461 | 1220 |
| ARC-AGI-2 | 61.3% | — |
| Kagi LLM Benchmark | 53.5% | — |
| NYT Connections (extended) | 91.3% | — |
| ARC-AGI-1 | 90.5% | — |
| CritPt | 18% | — |
| Chess Puzzles | 47% | — |
| Mystery Game Puzzles | 43% | — |
| DTBench | 93.9% | — |
| LMCA | 45.5% | — |
| Surface Evolver Bench | 40% | — |
| Epoch Capabilities Index | 155.31 | — |
| ForecastBench | 56.1 | — |
Math DeepSeek V4 Pro leads
DeepSeek V4 Pro: 64.8 (#30), Tulu 3 (Tülu 3) 70B: 14.2 (#303)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| OTIS Mock AIME 2024-2025 | 98.6% | 4.4% |
| LMArena Math | 1455 | 1242 |
| FrontierMath (Tiers 1-3) | 64.6% | — |
| FrontierMath Tier 4 | 26.8% | — |
| MathArena Final-Answer Competitions | 76.6% | — |
| ProofBench | 50% | — |
| MATH Level 5 | — | 42.7% |
Knowledge DeepSeek V4 Pro leads
DeepSeek V4 Pro: 59.5 (#31), Tulu 3 (Tülu 3) 70B: 25.0 (#264)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| GPQA Diamond | 91.7% | 46.3% |
| SimpleQA Verified | 52.9% | — |
| Vectara Hallucination Rate | 8.6% | — |
| LMArena Expert | 1464 | — |
Multilingual DeepSeek V4 Pro leads
DeepSeek V4 Pro: 54.4 (#45), Tulu 3 (Tülu 3) 70B: 39.9 (#222)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| LMArena Non-English | 1439 | 1236 |
| LMArena Chinese | 1486 | 1249 |
| LMArena Russian | 1453 | 1246 |
| LMArena French | 1472 | — |
| LMArena German | 1458 | — |
| LMArena Japanese | 1445 | — |
| LMArena Korean | 1447 | — |
| LMArena Spanish | 1458 | — |
Instruction Following DeepSeek V4 Pro leads
DeepSeek V4 Pro: 76.1 (#47), Tulu 3 (Tülu 3) 70B: 64.8 (#227)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| LMArena Instruction Following | 1448 | 1233 |
Long Context DeepSeek V4 Pro leads
DeepSeek V4 Pro: 45.0 (#51), Tulu 3 (Tülu 3) 70B: 37.1 (#222)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| LMArena Longer Query | 1458 | 1224 |
| CL-bench Life | 13.5% | — |
Writing & Preference DeepSeek V4 Pro leads
DeepSeek V4 Pro: 65.5 (#46), Tulu 3 (Tülu 3) 70B: 45.6 (#223)
| Benchmark | DeepSeek V4 Pro | Tulu 3 (Tülu 3) 70B |
|---|---|---|
| LMArena Text | 1451 | 1256 |
| LMArena Creative Writing | 1446 | 1231 |
| LMArena Multi-Turn | 1467 | 1252 |
| EQ-Bench Creative Writing | 1553 | — |
| EQ-Bench 4 | 1166 | — |
Frequently asked questions
Is DeepSeek V4 Pro better than Tulu 3 (Tülu 3) 70B?
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 33.0 on the Noometry Index.
Is DeepSeek V4 Pro or Tulu 3 (Tülu 3) 70B better for coding?
DeepSeek V4 Pro scores higher on coding benchmarks: 52.4 versus 36.0 in the Noometry coding category.
How many benchmarks do DeepSeek V4 Pro and Tulu 3 (Tülu 3) 70B share?
13 benchmarks have published results for both models. DeepSeek V4 Pro has 48 scored results on Noometry and Tulu 3 (Tülu 3) 70B has 14.