Search Intelligence Score
64.6
+33.9 lift from search
050100
- Lift from search
- +33.9
- 30.6 without search
- Cost per 1K tasks
- $93.2
- Below the median score of 66.0, so not ranked on cost
Accuracy by benchmark
Accuracy on each eval, with search (solid) and without (light).
- DSQA83.8 vs 41.0
- HLE39.0 vs 38.0
- WISER71.0 vs 13.0
Cost per 1K tasks
How the cost splits between model inference, search calls, and page extraction.
- Inference
- $79.5
- 85%
- Search
- $4.03
- 4%
- Extract
- $9.71
- 10%
Compared with
Other models from the same provider alongside models that score closest.
- Hy4 Preview64.6Cost $93.2 / Lift +33.9
- Hunyuan 358.2Cost $28.4 / Lift +35.6
- Hy475.4Cost $94.4 / Lift +45.3
- Mistral Large 464.4Cost $136 / Lift +43.0
- MiMo v2.6 Pro64.4Cost $76.5 / Lift +34.7
| Model | Score | Lift | Cost per 1K tasks |
|---|---|---|---|
| Hy4 Preview | 64.6 | +33.9 | $93.2 |
| Hunyuan 3Hunyuan 3 | 58.2 | +35.6 | $28.4 |
| Hy4Hy4 | 75.4 | +45.3 | $94.4 |
| Mistral Large 4Mistral Large 4.0 | 64.4 | +43.0 | $136 |
| MiMo v2.6 ProMiMo v2.6 Pro | 64.4 | +34.7 | $76.5 |
Search Intelligence Score averages accuracy on DSQA, HLE, and WISER with equal weight; cost is averaged the same way, with cost shown per 1,000 tasks. Evaluated . Leaderboard last updated . Full methodology.
# Hy4 Preview / Search Capability Leaderboard
- Search Intelligence Score: 64.6 (#31 in Search Intelligence)
- Without search: 30.6 / Lift from search +33.9
- Cost per 1K tasks: $93.2 (below the median score of 66.0, so not ranked on cost)
- Evaluated:
- Lab: Tencent / Weights: Open weights / Model id: tencent/hy4-preview
## Accuracy by benchmark
| Suite | With search | Without search |
|---|---|---|
| DSQA | 83.8 | 41.0 |
| HLE | 39.0 | 38.0 |
| WISER | 71.0 | 13.0 |
## Compared with
| Model | Score | Lift | Cost per 1K tasks |
|---|---|---|---|
| Hy4 Preview | 64.6 | +33.9 | $93.2 |
| Hunyuan 3 | 58.2 | +35.6 | $28.4 |
| Hy4 | 75.4 | +45.3 | $94.4 |
| Mistral Large 4.0 | 64.4 | +43.0 | $136 |
| MiMo v2.6 Pro | 64.4 | +34.7 | $76.5 |
Leaderboard last updated . Methodology: /leaderboard#methodology