--- license: apache-2.0 base_model: Qwen/Qwen3-Reranker-0.6B library_name: gguf pipeline_tag: text-ranking tags: - gguf - reranker - deepkit-code --- # Deepkit Code — Qwen3 Reranker 0.6B GGUF Q8_0 and f16 conversions of the official Qwen3-Reranker-0.6B model, prepared for [Deepkit Code](https://github.com/bartscheffer/deepkit-code). This repository hosts model artifacts; it does not contain the Deepkit Code executable. ## Provenance - Official model revision: `e61197ed45024b0ed8a2d74b80b4d909f1255473`. - Unmodified official llama.cpp converter: `e79e4bf660e19f2ad851e06c6913f7a8c5852621`. - Formats: Q8_0 and f16. Exact sizes and SHA-256 are in `artifact-manifest.json`. - Model license: Apache-2.0. The original model is developed by Qwen; this is a project-maintained conversion, not an official Qwen GGUF release. [Conversion scripts and dependency lock](https://github.com/bartscheffer/deepkit-code/tree/a5762f2/scripts/models) record the preparation workflow. All upstream files were verified against pinned Git-blob/LFS hashes before offline conversion. Repeating Q8_0 conversion produced byte-identical output in the recorded environment. ## Runtime contract These are rerankers, not embedding models. Both contain `cls.output.weight`, Rank pooling, classifier labels `[yes, no]`, and `tokenizer.chat_template.rerank`. Use the stored rerank template for the query/document pair. The recorded native Qwen3 Rank implementation already applies softmax: the first pooled value is P(yes). Do not apply a second softmax or L2-normalize the two probabilities. Both artifacts pass actual CPU three-pair checks: a relevant answer scores above 0.999 and an irrelevant answer below 0.00002. Classifier rows were also compared against the original yes/no token weights. Details are in the manifest. These are artifact smoke tests, not a retrieval benchmark or a throughput guarantee. Deepkit Code's production search integration is still in development.