A smaller version of Google's SigLIP 2 ViT-B/16 (WebLI) in ONNX, with the weights kept as 8-bit numbers in memory as well, not only in the file: the matrix multiplications run on 8-bit integers (ONNX Runtime's MatMulNBits). It is the sibling of vlk74/ViT-B-16-SigLIP2-cs-en-8bit, which unpacks its weights to 32-bit floats when it loads. The two give practically the same search results; this one takes about 500 MB less memory with both encoders loaded, and is slower on x86 processors. It is published to be measured on ARM phones, which have dedicated kernels for it.
This is a modified model and not an official release by Google, the OpenCLIP authors or the Immich project.
Starting point: the ONNX export immich-app/ViT-B-16-SigLIP2__webli. Changed on 1 November 2026:
| File | Change |
|---|---|
textual/model.onnx |
The word-piece table was cut from 256,000 to 82,948 rows and stored as 8-bit integers with one scale per row. The MatMul layers became MatMulNBits (8 bits, symmetric, one scale per block of 32 numbers, accuracy_level 4: the input is rounded to 8 bits per block too). |
textual/tokenizer.json |
Only the kept word pieces, renumbered, and the merges between them. |
visual/model.onnx |
The MatMul layers became MatMulNBits, as above. |
config.json, visual/preprocess_cfg.json, textual/tokenizer_config.json and textual/special_tokens_map.json are unchanged.
The kept word pieces are the ones used by 200,000 sentences per language (news and Wikipedia, from the Leipzig Corpora Collection), a small set of English and Czech image-tagging prompts and short phrases, every special token, the 256 byte pieces any other text falls back to, and the pieces those are merged from. The sentences themselves are not included here.
Made with ONNX Runtime 1.31's MatMulNBitsQuantizer (bits=8, block_size=32, is_symmetric=True, accuracy_level=4).
The inputs, the outputs and the file layout are those of the Immich export.
MatMulNBits is in its com.microsoft domain); 1.31 was used. No session option is needed.textual/tokenizer.json with the tokenizers library. tokenizer_config.json is the upstream file: its added_tokens_decoder still has upstream's numbers, 32 of which no longer match, so don't load this folder with AutoTokenizer.config.json says (canonicalize: lower-case, no ASCII punctuation), and pad or cut it to 64 tokens.200 photos from Crossmodal-3600 (Google, CC BY 4.0), each described by people in each language. Every description searches all the photos, and its own photo should come first.
| Czech: original | Czech: this model | English: original | English: this model | |
|---|---|---|---|---|
| Own photo comes 1st | 73.0 % | 73.0 % | 84.0 % | 84.0 % |
| In the top 5 | 93.0 % | 92.5 % | 99.0 % | 99.0 % |
| In the top 10 | 97.0 % | 96.5 % | 99.5 % | 99.5 % |
| Top 20 results the same as the original's | 96 % | 97 % |
Vectors against the original's (cosine):
| Average | Lowest | |
|---|---|---|
| Photos | 0.9997 | 0.9968 |
| Czech descriptions | 0.9989 | 0.8973 |
| English descriptions | 0.9990 | 0.9232 |
The lowest text values come from words whose word pieces were cut.
| Original | 8-bit weights (sibling) | This model | |
|---|---|---|---|
visual/model.onnx |
372 MB | 96 MB | 107 MB |
textual/model.onnx |
1,129 MB | 174 MB | 183 MB |
| Whole model | 1,536 MB | 276 MB | 296 MB |
| One photo (Intel i5-5200U, 2 threads) | 364 ms | 397 ms | 668 ms |
| One text (Intel i5-5200U, 2 threads) | 126 ms | 130 ms | 207 ms |
Tested on x86-64 Linux with ONNX Runtime 1.31. Not yet measured on ARM.
Apache License 2.0; see LICENSE. The changes above are stated as section 4(b) of the licence requires.
textual/model.onnx 183 MBvisual/model.onnx 107 MBtimm/ViT-B-16-SigLIP2