Benchmarks / Model
DeepSeek V4 Pro
Results for DeepSeek V4 Pro by DeepSeek across JumalAIta Intelligence benchmarks, with the configuration and run details of every evaluation.
01 / Result
Creative Reimplementation Benchmark
- Benchmark version
- v1.0
- Generation
- Single shot
- Reasoning effort
- High
- Harness
- DeepSeek Harness (dsh) 0.1.5-rc.3
- Run date
- 2026-09-26
- Run environment
- Reference container (see the protocol)
- Run time
- 13 min 11 s
- Tool calls
- 55
- Cost
- Not yet published
- Tokens
- Not yet published
- Run note
- DeepSeek V4 Pro does not accept image input, so it never saw the reference capture. Its own words in the run transcript: "I can't see the video frames because the model doesn't support image input." It worked from the soundtrack and the average colours of the frames instead.
- Evaluators
- The original authors of the reference demo (members of Jumalauta)
- Recording length
- 2 min 55 s
- Recording published
- 2026-09-26
Evaluators' notes
- Rank 12
DeepSeek V4 Pro
According to the original authors, the DeepSeek V4 Pro effects are better than Claude Haiku 4.5's, but the demo shows no trace of the spirit of the original production.
02 / Summary
Benchmark results
| Benchmark | Version | Rank | Evaluation | Configuration | Status | Result |
|---|---|---|---|---|---|---|
| Creative Reimplementation Benchmark | v1.0 | 12 | Expert-ranked | Single shot · High | Final | Details of DeepSeek V4 Pro in Creative Reimplementation Benchmark |
