Jev vs. Laya
Same three question types, different ownership. Figures are the vendors' unless marked, and head-to-head numbers weren't always measured in the same run.
| Jev | Laya | |
|---|---|---|
| Maker | TypeSafe AI | Convai Innovations |
| Released | 15 Sep 2026 | 18 Sep 2026 |
| Weights | Closed, hosted API only | Open, Apache 2.0 |
| Price | $0.042 / M input tokens, output free | $0, plus your hardware |
| Latency | 70–500 ms end to end | 32.8 ms on a T4; 7.2 ms per question batched |
| Zero-shot quality | Strong | Near chance on the base checkpoint (0.362 vs 0.318 random) |
| Calibration (ECE, lower is better) | 0.144 | 0.213 |
| Large label sets (Banking77) | 0.870 on 72 labels | 0.425 on 77 labels |
| Fine-tuning | Not available | Yes |
| Languages | Not specified at launch | 100+ via the multilingual checkpoint |
| Data leaves your server | Yes | No |
| Best fit | No labeled data yet, many labels, no infrastructure | Private data, high volume, labeled examples to fine-tune on |
Laya's “6–8× faster” compares local GPU inference with Jev's network round trip. Both numbers can be true without the comparison being like-for-like. See the entry.