Liquid AI · Open weights, local
d1-3B by Liquid AI scores 78.4 on the System One Index (10th of 24). Calibration error 2.4%.
| Dataset | Accuracy | Rank | Best on the board | Calibration error |
|---|---|---|---|---|
| AG News English | 88.5 | 14th of 24 | 92.5 | 7.1% |
| SST-2 English | 95.5 | 4th of 24 | 97.0 | 4.4% |
| SST-5 English | 56.0 | 8th of 24 | 62.5 | 5.7% |
| DAIR Emotion English | 63.5 | 5th of 24 | 91.5 | 13.8% |
| Banking77 English | 88.0 | 8th of 24 | 98.5 | 3.1% |
| BoolQ English | 85.0 | 17th of 24 | 92.5 | 6.1% |
| XNLI English English | 86.0 | 12th of 24 | 93.3 | 3.1% |
| Typed decisions English | 64.8 | 14th of 24 | 79.0 | 3.1% |
| XNLI French French | 78.7 | 13th of 24 | 90.7 | 7.0% |
| XNLI Arabic Arabic | 73.3 | 9th of 26 | 89.3 | 7.5% |
| MASSIVE scenario Arabic | 71.5 | 8th of 26 | 89.5 | 7.7% |
| MASSIVE intent Arabic | 83.5 | 11th of 26 | 91.5 | 6.1% |
| OSACT4 offensive Arabic | 84.5 | 15th of 26 | 95.5 | 3.8% |
| Tweet sentiment Arabic | 63.5 | 12th of 26 | 80.0 | 11.2% |
| MIRACL rerank Arabic | 82.5 | 14th of 26 | 90.5 | 6.5% |
| MIRACL pairwise Arabic | 89.0 | 7th of 26 | 91.0 | 5.2% |
LiquidAI/d1-3BOpen-weight 3B version of d1, text only (no mmproj). LFM Open License. Served with 4,096-token batches and one slot; with the 8,192-token setting the other GGUFs used, its server died, most likely out of memory.
Ranks count the models charted on the leaderboard. Every model gets the same cases and the same request, under the benchmark rules.