~/system-one-bench / compare/fastino-glide-vs-jared-palmer-kev-4b
Head to head
GLiDE scores 84.1 and Kev 4B 76.8 on the System One Index. GLiDE is more accurate on 16 of 16 text datasets, Kev 4B on 0. Same cases, same request, same scorer.
| GLiDE | Kev 4B | |
|---|---|---|
| System One Index | 84.1 | 76.8 |
| English | 83.0 | 78.3 |
| Arabic | 84.8 | 74.8 |
| Vision Index | 82.5 | – |
| Calibration error | 3.6% | 5.1% |
| Median latency | – | 552 ms |
| Cost per 1,000 calls | $0.017 | $0.010 |
| GLiDE | Kev 4B | |
|---|---|---|
| AG News English | 92.0 | 90.5 |
| SST-2 English | 96.0 | 92.5 |
| SST-5 English | 61.5 | 52.0 |
| DAIR Emotion English | 61.5 | 54.5 |
| Banking77 English | 91.5 | 88.0 |
| BoolQ English | 92.0 | 91.5 |
| XNLI English English | 94.7 | 90.7 |
| Typed decisions English | 75.0 | 66.3 |
| XNLI French French | 88.0 | 79.3 |
| XNLI Arabic Arabic | 83.3 | 72.7 |
| MASSIVE scenario Arabic | 73.5 | 57.0 |
| MASSIVE intent Arabic | 91.0 | 76.5 |
| OSACT4 offensive Arabic | 95.0 | 88.5 |
| Tweet sentiment Arabic | 73.5 | 63.0 |
| MIRACL rerank Arabic | 87.0 | 79.5 |
| MIRACL pairwise Arabic | 90.5 | 86.5 |
| GLiDE | Kev 4B | |
|---|---|---|
| SST-2, as an image | 94.0 | – |
| Arabic sentiment, as an image | 51.5 | – |
| Oxford-IIIT Pet | 84.5 | – |
| VQAv2 yes/no | 89.0 | – |
| CountBench | 93.5 | – |
Model pages: GLiDE · Kev 4B. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.