ONNX conversion of gliner-community/glinersmall-v2.5, dynamically quantised to INT8, packaged as a self-contained bundle for offline NER. This is a re-serialisation, not a fine-tune: the weights are the upstream ones. Only the format (PyTorch → ONNX) and the precision (fp32 → INT8) are ours. The fp32 reference graph (model.onnx, sha256 5245733ccb2b75072cce0b4bbb14424988f92f9daf775d97bdf0de74be28df63) is not shipped — it is only needed to reproduce the INT8 graph. Its hash is recorded in NOTICE. The graph has six inputs, fed per span-encoded prompt: spanmask is bool (not int64) — the graph declares tensor(bool). The prompt follows the GLiNER label format, using the model's own special tokens…
SAVRN Model Hub · Models by Task
Token Classification Models
61 open-weight token classification models in the SAVRN Model Hub, with OpenMed, D and Microsoft publishing the most.
61Models
25Publishers
14M to 1.4BParameter range
6Licenses
SAVRN's Take
Pull names, places, dates and personal identifiers out of a document, restore lost punctuation, or de-identify a radiology report, and you are doing token classification. What people run is small: Bayerische Staatsbibliothek's bert-large-cased-finetuned-conll03-english drew 3,460,219 downloads a month, Wilson Wongso's Indonesian tagger 2,601,173, Stanford AIMI's de-identifier 1,442,058 and bert-base-NER 1,429,534. OpenMed published 31 of the 61 entries, and its privacy-filter-multilingual, the largest at 1.4B parameters, tags 54 categories of personal data in 16 languages.
None of this needs a cluster. Parameters run from 14M for Leo Kang's KoELECTRA-small-v3-modu-ner to 1.4B for OpenMed's privacy-filter-multilingual, and memory follows: 0.8 GB at 16-bit for the download leader, 0.3 GB for bert-base-NER, 3.4 GB at 16-bit or 0.8 GB at 4-bit for the largest. The cheapest host the Index lists for every model it has sized is one MI300X at $1.85 an hour, more card than any of these can fill, so stack several taggers on one GPU. Context is the limit: the BERT and RoBERTa fine-tunes take 512 or 514 tokens per pass, so a long contract gets chunked, while the 1.4B model takes 131,072.
Licensing is mostly permissive, 42 of 61 Apache 2.0 and 10 MIT, but the edges bite. Five entries state no license, the download leader among them, so a commercial deployment of it has nothing to show counsel. Babelscape's wikineural-multilingual-ner, 677,069 downloads a month, is CC BY-NC-SA 4.0 and belongs in no paid product. Check license, label set and language before committing; many of these were tuned for one language only.
Most Downloaded
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| bert-large-cased-finetuned-conll03-english | Bayerische Staatsbibliothek | 334M | Not stated | 3.5M | 1x MI300X, $1.85/hr |
| indonesian-roberta-base-posp-tagger | Wilson Wongso | 124M | mit | 2.6M | 1x MI300X, $1.85/hr |
| stanford-deidentifier-base | Stanford AIMI | — | mit | 1.4M | — |
| bert-base-NER | D | 108M | mit | 1.4M | 1x MI300X, $1.85/hr |
| bert-portuguese-ner | Luís Filipe Cunha | — | mit | 1M | — |
| privacy-filter-nemotron-GGUF | LocalAI-io | — | apache-2.0 | 951.7k | — |
| punctuate-all | KREDOR | — | mit | 815k | — |
| wikineural-multilingual-ner | Babelscape | 177M | cc-by-nc-sa-4.0 | 677.1k | 1x MI300X, $1.85/hr |
| sat-3l-sm | Segment any Text | 214M | mit | 515.8k | 1x MI300X, $1.85/hr |
| KoELECTRA-small-v3-modu-ner | Leo Kang | 14M | Not stated | 401.8k | 1x MI300X, $1.85/hr |
Licenses
| License | Models | Commercial use |
|---|---|---|
| apache-2.0 | 42 | Yes |
| mit | 10 | Yes |
| not stated | 5 | Not stated |
| afl-3.0 | 2 | Read the license |
| cc-by-nc-sa-4.0 | 1 | Not without separate permission |
| cc-by-nc-nd-4.0 | 1 | Read the license |
Who Publishes Them
| Publisher | Models |
|---|---|
| OpenMed | 31 |
| D | 3 |
| Microsoft | 2 |
| LocalAI-io | 2 |
| Flair | 2 |
| David Adelani | 2 |
All 61 Models, Page 2 of 2
Questions
Which Token classification models are most downloaded?
By monthly downloads reported by the Hugging Face Hub: .