GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ZakariaAlMoktar
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ZakariaAlMoktar
Model Tree
Qwen/Qwen3-Reranker-0.6B
Input Modalities
Output Modalities
Supported Functionality
| Property | Value |
|---|---|
| Base model | Qwen/Qwen3-Reranker-0.6B |
| Base revision | e61197ed45024b0ed8a2d74b80b4d909f1255473 |
| Released checkpoint | checkpoint-1235 |
| Adapter | LoRA r=16, alpha=32, dropout=0.05 |
| Target modules | q/k/v/o attention projections |
| Maximum sequence length | 256 tokens |
| Training objective | listwise LambdaLoss |
| Training hardware | Single NVIDIA A40, 48 GB VRAM |
python
import torch.nn as nnfrom sentence_transformers import CrossEncodermodel = CrossEncoder("ZakariaAlMoktar/presto-qwen3-reranker-0.6b-arabic-ecommerce",activation_fn=nn.Identity(),device="cuda",)pairs = [("قهوة عربية", "قهوة عربية محمصة 500 جرام"),("قهوة عربية", "شاحن هاتف سريع"),]scores = model.predict(pairs)
The model returns raw cross-encoder scores. Calibrate scores within each query when combining them with retrieval or lexical channels.
Each list contains up to four relevant products and four labeled hard
negatives. The released checkpoint-1235 achieved provider-validation
nDCG@10 of 0.9750717060. The complete training recipe is available in the
Presto Arabic Search repository.
The adapter is designed to rerank a bounded candidate set. It is not a full-catalog retriever and should be applied after candidate generation.
The score is uncalibrated and is most useful as one relevance signal inside a multi-signal ranking system. Performance may shift on catalogs whose language, taxonomy, or title style differs from the training domain.
Apache-2.0. The adapter requires the Apache-2.0 Qwen base model.