Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models
Paper • 2506.05176 • Published • 86
Qwen3-Embedding-0.6B for sidekick (Core ML)
Qwen/Qwen3-Embedding-0.6B, a Qwen3 decoder embedding model (1024 dimensions), converted to Core ML by sidekick for Apple silicon's Neural Engine, GPU and CPU. This repository holds one build, fp16/, graded against the original model on every compute path it runs on.
Licence: Apache-2.0 · publishing tier 2a. Training data with non-commercial terms. Nothing to accept, but read this before you use the model:
- The data: MS MARCO (non-commercial research only).
- Whether such terms bind a model trained on the data is legally unsettled. You decide whether your use is permitted, and the publisher of this repository gives no assurance that it is.
- The model's own licence is unchanged: these files are distributed under Apache-2.0, as the original model is.
Why: Mostly trained on about 150M synthetic pairs generated by Qwen3-32B from an undisclosed corpus, and fine-tuned on labelled sets that include MS MARCO (non-commercial research only) and Natural Questions (CC BY-SA). Non-commercial training data: 2a, with its terms stated on the card. The training data below gives every source's licence. How sidekick decides what it publishes.
Download
| Build | Weights | Sequence buckets | Max tokens | Download | Shared weights | Runs on (grade) |
|---|---|---|---|---|---|---|
fp16/ |
16-bit float | 128, 256, 512, 1024, 2048 (2 programs each) | 2048 | 1.23 GB | Yes, one file for all buckets | Neural Engine (default) A, GPU A, CPU B |
sidekickd fetch qwen3-embedding-0.6b
Needs sidekick 0.16.0 or later. How fetch works · Installing without fetch · Calling the model · Why compiled programs
Fidelity
Graded against the original model in fp32 by sidekick's parity suite. What the grades mean.
| Compute path | Chip | OS | Grade | Worst cosine | Graded files |
|---|---|---|---|---|---|
| Neural Engine (default) | Apple M1 Max | macOS 27.0 | A | 0.99999 | This build |
| GPU | Apple M1 Max | macOS 27.0 | A | 0.999997 | This build |
| CPU | Apple M1 Max | macOS 27.0 | B | 0.99968 | This build |
This build: the grading report's file hashes match this repository's files. Unconfirmed: graded on a build of the same model that can't be matched to these files by hash.
- Neural Engine, Apple M1 Max: Long inputs (4 cases of 600–2,048 tokens, fixtures/parity/corpus_long.toml): A, worst cosine 0.99999
- GPU, Apple M1 Max: Long inputs (4 cases of 600–2,048 tokens, fixtures/parity/corpus_long.toml): A, worst cosine 0.999997
- CPU, Apple M1 Max: Served and graded up to 512 tokens on this path (its manifest's cpu_max_seq_len, below the CPU's cap; D33); 3 longer cases not run
Benchmark scores
The original model's published scores, not this build's. Every score, with its caveats.
| Benchmark | Score | Measure | Source |
|---|---|---|---|
| MTEB(eng, v2) | 70.47 | Mean over its tasks | MTEB results |
| MTEB(Multilingual, v2) | 64.34 | Mean over its tasks | MTEB results |
| MTEB(eng, v1) | 67.05 | Mean over its tasks | MTEB results |
sidekick also ran MTEB tasks through fp16/ itself: its scores beside the original's.
Training data
Training data: partly disclosed. Each dataset under its original source's licence:
| Dataset | Licence | Terms | Stage |
|---|---|---|---|
| Synthetic weakly supervised pairs (~150M) generated by Qwen3-32B from the Qwen3 pre-training corpus | unknown | Unknown | weakly supervised pre-training |
| Synthetic high-quality pairs (~12M, filtered) | unknown | Unknown | supervised fine-tune |
| MS MARCO | MS MARCO terms (non-commercial research only) | Non-commercial | supervised fine-tune |
| Natural Questions | CC-BY-SA-3.0 | Share-alike | supervised fine-tune |
| Other labelled sets (HotpotQA, NLI, DuReader, T2-Ranking, SimCLUE, MIRACL, MLDR, Mr.TyDi, Multi-CPR, CodeSearchNet, 'and others') | unverified | Unknown | supervised fine-tune |
Not confirmed from a primary source
- Full list of labelled datasets (paper says 'and others').
- Paper content read through a summarising fetch of the arXiv HTML.
Original model
| Model | Qwen/Qwen3-Embedding-0.6B |
| Revision | 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3 |
| Licence | Apache-2.0 |
| Licence file | LICENSE: the Apache License 2.0 as the Apache Software Foundation publishes it; the model's repository carries no licence file |
| Base model | Qwen/Qwen3-0.6B-Base (Apache-2.0) |
| Changes in this repository | Converted to Core ML in 16-bit float by tools/convert_qwen3_embedding.py, with a sidekick manifest added. Reproduce it |
Checksums
fp16/: ledger build qwen3-embedding-0.6b/coreml/fp16/2c652898a8f8, 52 files, 2 shared
sidekickd fetch checks every file against sidekick's checksum ledger; fp16.sha256 has the same list for shasum -a 256 -c.
| Built with | |
|---|---|
| Converter | tools/convert_qwen3_embedding.py at cf8cd3dc1d11 |
| coremltools | 9.0 |
| PyTorch | 2.13.0 |
| Python | 3.12.13 |
| OS | macOS 27.0 (26A428) |
| Xcode | 27.1 (27A9275) |
| Converted on | Apple M1 Max |
| Converted | 2026-10-23 |
| Shared weights | Yes: one weights file for every bucket |
| sha256 | Size | Path |
|---|---|---|
475c70ed6609f077bdb6be750311bcb0e853dbe9f26f707c85933124fe15807c |
12.5 kB | manifest.toml |
f62be9e6741f4566a0a5e999227112a622726ac84548991ebf38695bfba620bd |
243 B | model_1024.0.mlmodelc/analytics/coremldata.bin |
e5e09f9652fc8ab7a029882d2b5fbe084d971b044313a868c1cc28d97eb1edd7 |
412 B | model_1024.0.mlmodelc/coremldata.bin |
1afdb042e265f41e181340a546db33b81c2a9c250e83b10c80bd2ac2910f70c0 |
2.5 kB | model_1024.0.mlmodelc/metadata.json |
8cc3020ab920b841c62eb4d13a9a9791a9f6d9085f80802d3ac20269be39a422 |
207.6 kB | model_1024.0.mlmodelc/model.mil |
47e408fb5ffceb898899e904f7ccafc137b553c0ec35933dedc82648ece4e536 |
607.0 MB | model_1024.0.mlmodelc/weights/weight.binmodel_128.0.mlmodelc/weights/weight.binmodel_2048.0.mlmodelc/weights/weight.binmodel_256.0.mlmodelc/weights/weight.binmodel_512.0.mlmodelc/weights/weight.bin |
9e2e20167facd344104f3beaa81463ce997242fce04e867cfb21eee6f1359b2b |
243 B | model_1024.1.mlmodelc/analytics/coremldata.bin |
cc292084d1f55d49d038f2bb6d89b63a161bcdec4517c234d7ad181c8f1c13ed |
411 B | model_1024.1.mlmodelc/coremldata.bin |
de4c1df5b424e07b219db205d1d1c3698fdc7cd5038e73ed88f28923ad6c57bf |
2.5 kB | model_1024.1.mlmodelc/metadata.json |
c76fde59899c2b23fb014ee1b7178d3eb5dbb3342d917b82c0deb65766abbef8 |
433.1 kB | model_1024.1.mlmodelc/model.mil |
14564d436401146248ccf5a2cb6e6a0e4ebddac679b55d995d296fe71f21671a |
611.0 MB | model_1024.1.mlmodelc/weights/weight.binmodel_128.1.mlmodelc/weights/weight.binmodel_2048.1.mlmodelc/weights/weight.binmodel_256.1.mlmodelc/weights/weight.binmodel_512.1.mlmodelc/weights/weight.bin |
d205570b02f16ac2b09bf8b6ad71362cb586c7f3b6cc1a356af668e4c40bbaab |
243 B | model_128.0.mlmodelc/analytics/coremldata.bin |
e6c8c25b30a898852cadaa4e5854fc0cf4463b691bbc5c81467a4a54ce600cf1 |
412 B | model_128.0.mlmodelc/coremldata.bin |
f917311e02a7a5eb012322b6bca07e8889b9aed59d68a40439493c9d3e40e5ce |
2.5 kB | model_128.0.mlmodelc/metadata.json |
21db2146acbf737c641d1e24a5976a865a9d46b54a0284df11621d83ac8fc81f |
206.8 kB | model_128.0.mlmodelc/model.mil |
ff6b4e44fc8f6b678ecc648b263a4a00ac398936a352d566daaafe1a4711ff5c |
243 B | model_128.1.mlmodelc/analytics/coremldata.bin |
ab9f9eeb614aff718905455b0804242bead2662f6e3dae29975464a7b55af143 |
411 B | model_128.1.mlmodelc/coremldata.bin |
62d0e330aedec6f317db723b860c8f21d36ee7956b5cbf6e8b1cd39e93c5c661 |
2.5 kB | model_128.1.mlmodelc/metadata.json |
b98d74ad1644f9fbd1aeb606ba0dcb6c686ec9d2068d7160080baaf110cb7674 |
431.4 kB | model_128.1.mlmodelc/model.mil |
9fd1f76254863763d52fb335d8b1705f6c9206129c33c7fccc4ae7eb2fe729eb |
243 B | model_2048.0.mlmodelc/analytics/coremldata.bin |
6087936a73582c99921102f45203bb8a68e21d91901bed71945fe0eb95f9f7ed |
412 B | model_2048.0.mlmodelc/coremldata.bin |
d2f14f76116852ae6bacbd7a2f970330fa94ae4682b99d2c760b1ae601bef08a |
2.5 kB | model_2048.0.mlmodelc/metadata.json |
5b65e6c2fead703bd63185b9435cbc36a82448be2ced13cb26ec176d91a3c05e |
207.6 kB | model_2048.0.mlmodelc/model.mil |
a10ebfac8abd5397b05b9bb024eb9d0ee2e6577706a6e2bbc5b304ba45b973df |
243 B | model_2048.1.mlmodelc/analytics/coremldata.bin |
0d5bb237177d9b961b06890081f862e508cfaa3322d04aea51a5c33edbaac0e3 |
411 B | model_2048.1.mlmodelc/coremldata.bin |
36eccef2e8ac4381a50a0e1424386b149776b4e3ef52c2ff5de00b930c696499 |
2.5 kB | model_2048.1.mlmodelc/metadata.json |
f65593c08705f331ddf6d62b515c31c0ba5e1449cf73f784591da15e61326b19 |
433.1 kB | model_2048.1.mlmodelc/model.mil |
64277ca26265778717d29c110dacdbda09546ae86144e6db55cf83fc62c97ac4 |
243 B | model_256.0.mlmodelc/analytics/coremldata.bin |
d851c4d55c49dbaca5e4bfcc92faebc3a7dce8b38fec30acd65eb08429347c44 |
412 B | model_256.0.mlmodelc/coremldata.bin |
eae51638073f5880f70cd450854ef881f3103e8265f249d1174d487147075734 |
2.5 kB | model_256.0.mlmodelc/metadata.json |
c1555252b6c1f37851552fd87b01005fc8481ee335a66f9a0b5f6c5e46df796b |
206.8 kB | model_256.0.mlmodelc/model.mil |
54fedbd8bd9e353ca50f48787445245eca0811ecb2f9eba0e734b58228483cca |
243 B | model_256.1.mlmodelc/analytics/coremldata.bin |
a559c85ec1d68562e6a3a0dc0b11d56dad34d834ce935ecc8dc592d5a5e8e39b |
411 B | model_256.1.mlmodelc/coremldata.bin |
d112052d29dd921bc9ce0e3a5131c2aff60d4c6372c8212d4cf79c85f14398e7 |
2.5 kB | model_256.1.mlmodelc/metadata.json |
920f577589c78f6de0871542dee7b1f05021ec8649eab9a960a7dc2ce4128087 |
431.4 kB | model_256.1.mlmodelc/model.mil |
31e2d30b9876a243438a994506b4db829faa4a6f8e0ad1bdaa658d4b8ca264fd |
243 B | model_512.0.mlmodelc/analytics/coremldata.bin |
dcfcf2ed63865c7257df3ed5329de1642e5e372da433b0fc86a3c43e1fc80f04 |
412 B | model_512.0.mlmodelc/coremldata.bin |
1aff4a43ffa0a190b454d10409ee2c36e6a96afcbdd0c46263c852df69f18467 |
2.5 kB | model_512.0.mlmodelc/metadata.json |
18850c69ccc910b0095ef7736ca92ba8e720fc96d4b680029011631b9092e12c |
206.8 kB | model_512.0.mlmodelc/model.mil |
8f0568892ed099d3ba6fd1046eeb6dfeccf6bc7fd9e3aeb76cf14f35c9d23b84 |
243 B | model_512.1.mlmodelc/analytics/coremldata.bin |
33b548cf5be06c7460978c66186369c18708c7611c5f185cf9a0e3c64f1b920d |
411 B | model_512.1.mlmodelc/coremldata.bin |
e853e90d6785f088d6a77598b8c286e35a5d00a3a71aff417962d9160d1780de |
2.5 kB | model_512.1.mlmodelc/metadata.json |
97c006a023471d6aa146689e234a2800b76aeba6f848ced2d09972a1d985a1ea |
431.4 kB | model_512.1.mlmodelc/model.mil |
def76fb086971c7867b829c23a26261e38d9d74e02139253b38aeb9df8b4b50a |
11.4 MB | tokenizer.json |
Generated by tools/model_card.py from sidekick's committed data. Report a problem at https://github.com/ironblock/sidekick/issues.