Head to head
Clef scores 81.6 and Tev1 4B Experimental 71.5 on the System One Index. Clef is more accurate on 12 of 16 text datasets, Tev1 4B Experimental on 3, and 1 are tied. Same cases, same request, same scorer.
| Clef | Tev1 4B Experimental | |
|---|---|---|
| System One Index | 81.6 | 71.5 |
| English | 80.6 | 66.4 |
| Arabic | 83.0 | 75.8 |
| Vision Index | 78.1 | – |
| Calibration error | 2.2% | 5.3% |
| Median latency | 310 ms | 232 ms |
| Cost per 1,000 calls | $0.116 | $0.019 |
| Clef | Tev1 4B Experimental | |
|---|---|---|
| AG News English | 90.5 | 91.5 |
| SST-2 English | 97.0 | 94.0 |
| SST-5 English | 51.5 | 43.0 |
| DAIR Emotion English | 58.5 | 56.5 |
| Banking77 English | 95.5 | 1.0 |
| BoolQ English | 91.0 | 91.0 |
| XNLI English English | 88.7 | 87.3 |
| Typed decisions English | 72.2 | 67.0 |
| XNLI French French | 80.0 | 82.7 |
| XNLI Arabic Arabic | 78.0 | 76.0 |
| MASSIVE scenario Arabic | 88.0 | 56.5 |
| MASSIVE intent Arabic | 89.5 | 70.0 |
| OSACT4 offensive Arabic | 89.5 | 84.5 |
| Tweet sentiment Arabic | 58.0 | 70.5 |
| MIRACL rerank Arabic | 89.0 | 85.5 |
| MIRACL pairwise Arabic | 89.0 | 87.5 |
| Clef | Tev1 4B Experimental | |
|---|---|---|
| SST-2, as an image | 95.5 | – |
| Arabic sentiment, as an image | 35.0 | – |
| Oxford-IIIT Pet | 82.5 | – |
| VQAv2 yes/no | 87.5 | – |
| CountBench | 90.0 | – |
Model pages: Clef · Tev1 4B Experimental. Latency for API models is the median per call through OpenRouter, timed from one machine.