Qwen3-Embedding-0.6B for Core AI: asset, tokenizer files, config

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+embed-qwen3-0.6b-float32.aimodel/main.mlirb filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,47 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3-Embedding-0.6B
+tags:
+  - text-embeddings
+  - sentence-similarity
+  - retrieval
+  - core-ai
+  - apple-silicon
+  - macos
+library_name: swift-text-embedder
+---
+
+# embed-qwen3 — Qwen3-Embedding-0.6B for Core AI (macOS 27)
+
+Alibaba's [Qwen3-Embedding-0.6B](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B) (Apache 2.0)
+exported to an Apple Core AI `.aimodel` for on-device text embeddings and semantic search. Read by
+[swift-text-embedder](https://github.com/arraypress/swift-text-embedder) and the
+[`embed`](https://github.com/arraypress/swift-embed-cli) command-line tool.
+
+## Files
+
+| Path | What |
+|---|---|
+| `embed-qwen3-0.6b-float32.aimodel/` | Entry points `t128` and `t512`: token ids `[1, L]` + attention mask `[1, L]` (int32, LEFT-padded with `<|endoftext|>` = 151643) → the last position's hidden state `[1, 1024]`. Float32, 2.4 GB. |
+| `embed-support/vocab.json`, `merges.txt` | The checkpoint's own Qwen2 byte-level BPE files. |
+| `embed-support/config.json` | The pad / end-of-text id, the exported lengths, the dimension. |
+
+The host tokenises (NFC, Qwen2's pre-tokeniser, end-of-text appended), pads on the left, picks the
+shortest length that holds the text, and L2-normalises the output. Queries are embedded as
+`Instruct: <task>\nQuery: <text>`, documents as they are, per the model card.
+
+## Fidelity
+
+Against `transformers` 5.17 on the same inputs: tokenizer ids and masks identical over 13 texts at
+both lengths (Unicode, code, newlines included); embeddings 125–126 dB PSNR, cosine ≥ 0.999999.
+The fixed-length left padding changes an embedding by at most 3e-7 against the unpadded model.
+
+## Use
+
+```sh
+hf download arraypress/embed-qwen3 --local-dir models
+embed model install models
+embed index ~/Notes && embed search "how do I rotate the API key"
+```
+
+Requires macOS 27 (Core AI) on Apple silicon. Converted with coreai-torch 0.4.2 / torch 2.13.

embed-qwen3-0.6b-float32.aimodel/main.hash

@@ -0,0 +1 @@
+�)X	�Q�%��
u�`�����b��^<?e\
~�

embed-qwen3-0.6b-float32.aimodel/main.mlirb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2852958098d519625fcb70175956080aab195f362e418bb5e3c3f655c017efe
+size 2384530505

embed-qwen3-0.6b-float32.aimodel/metadata.json

@@ -0,0 +1,8 @@
+{
+  "assetVersion" : "2.0",
+  "license" : "Apache-2.0",
+  "author" : "Alibaba Qwen team — Qwen3-Embedding-0.6B; Core AI export by TextEmbedder",
+  "producer" : "coreai-core 1.0.0b2",
+  "description" : "Qwen3-Embedding-0.6B: entry points t128, t512 take left-padded token ids [1,L] + mask [1,L] (int32, pad 151643) and return the last position's hidden state [1,1024]; normalise in the host. Queries: 'Instruct: <task>\\nQuery: <text>'.",
+  "creationDate" : "20260924T123107Z"
+}

embed-support/config.json

@@ -0,0 +1,9 @@
+{
+ "model": "Qwen/Qwen3-Embedding-0.6B",
+ "lengths": [
+  128,
+  512
+ ],
+ "pad": 151643,
+ "dimension": 1024
+}