KaLM-Jev
A training-free Choice, Score and Noul service from the KaLM-Embedding team at HIT Shenzhen. Each option, level or yes/no criterion is scored as a document by the frozen KaLM-Reranker-V1-R2 rerankers (Nano, Small, Large), giving a probability per option with no generated text.
KaLM-Jev adds no trained weights: fixed instruction adapters drive KaLM-Embedding's own KaLM-Reranker-V1-R2 checkpoints (T5Gemma-2 based, Apache-2.0), one model per local HTTP process, so there are no files to pull here. The KaLM-Jev code carries no licence. Built by Xinping Zhao, first author of KaLM-Reranker-V1. The authors say the outputs are not calibrated probabilities and that their results do not establish Jev-equivalent quality. On their own 36-case smoke set, Choice accuracy is 67%, 92% and 83% for Nano, Small and Large, but Noul without supplied criteria answered yes on every case. Validated on an H100 MIG slice only; a demo Space runs Small.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Base model | KaLM-Embedding/KaLM-Reranker-V1-Nano-R2 |
| Maker | HIT-TMG (Lychee Team) |
| Released | 2026-09-21 |
| License | No licence on the KaLM-Jev code (treated as non-commercial); KaLM-Reranker-V1-R2 weights Apache-2.0 |
| Reported accuracy | 52.8% |
| Reported latency | about 52 ms p50 for a 3-option Choice on Nano (30 ms with a warm document cache), batch 8, H100 MIG slice |
Read more
This page was opened by System One for HIT-TMG (Lychee Team), who can claim the organisation and take it over at any time.