Hugging Face's logo

VAGOsolutions
/
SauerkrautLM-GLiNEREL-72m

Token Classification
Safetensors
glinerel-fast
extractor
glinerel
sauerkrautlm
named-entity-recognition
relation-extraction
information-extraction
pii-detection
pruning
knowledge-distillation
Model card Files
xet
 Community

SauerkrautLM-GLiNEREL-72m

Joint named-entity recognition, PII detection and relation extraction in one encoder pass. It turns a 33-page PDF into a knowledge graph in 0.34 s on a GPU, where a 4B generative model needs 17.6 s for a single paragraph. Labels are free text, chosen at inference time. English, German, French, Italian, Spanish. Try it in the demo →

Parameters 72.5M (146 MB)
Origin pruned + distilled from SauerkrautLM-GLiNEREL-314m: 23% of the parameters, 93% of the score
Overall score (4 relation benchmarks) 61.8
Latency per text 2.67 ms NVIDIA H100 (fp16) · 26.7 ms CPU (Apple M3 Pro, 3.8× faster than the reference)
License LFM Open License v1.0
Demo SauerkrautLM-GLiNEREL Interactive Knowledge Graph: PDF → knowledge graph, text extraction, benchmarks

A 33-page PDF becomes an interactive knowledge graph in the demo: SauerkrautLM-GLiNEREL-72m on the German Wikipedia article Siemens in 0.34 s on ZeroGPU and 2.67 ms per text on an NVIDIA H100; a 4B generative model needs 17.6 s for one paragraph.

Try it: SauerkrautLM-GLiNEREL Interactive Knowledge Graph: upload a PDF, explore the graph and export it as JSON, CSV, GraphML or Cypher (Neo4j).

The SauerkrautLM-GLiNEREL family on the main benchmarks: overall score, NER and relation F1, PII, multilingual NER, German relations, GPU latency on an NVIDIA H100, CPU latency on an Apple M3 Pro and model size.

Quickstart: glinerel-fast

glinerel-fast is our new inference library for SauerkrautLM-GLiNEREL, written for CPUs: about 1,600 lines of plain PyTorch, no transformers; also runs on NVIDIA and Apple GPUs.

pip install -U glinerel-fast     # private repo: hf auth login

from glinerel_fast import GLiNERel

model = GLiNERel.from_pretrained("VAGOsolutions/SauerkrautLM-GLiNEREL-72m")   # device="cuda", precision="bf16" on GPUs
result = model.extract(
    "Marie Curie was born in Warsaw. She married Pierre Curie, a professor at the University of Paris.",
    entities=["person", "location", "organization"],
    relations=["born in", "spouse", "works for"],
)
# {"entities": [{label, text, start, end, score}, ...], "relations": [{label, head, tail, score}, ...]}

model.redact("Call Laura Bianchi at +39 347 555 0192.", ["first name", "last name", "phone number"])
# 'Call [FIRST_NAME] [LAST_NAME] at [PHONE_NUMBER].'

What makes it fast

  • Packed batches: the tokens of all texts are concatenated without padding; only attention sees a padded view.
  • Fused matrices: q/k/v and the SwiGLU gate/up projections are one matrix product each; the LFM2 short convolutions run as three shifted multiply-adds that never cross text boundaries.
  • No 4-D intermediate in span scoring: the FiLM layer is rewritten algebraically, so the texts × labels × spans × width tensor is never built.
  • Cache-sized CPU batches: texts sorted by length, at most ~2,048 tokens per batch so activations stay in the CPU cache; tokenized words and label prompts are cached.
  • Inference-only weights: training heads are dropped at load time, no transformers model is built.
  • GPU (0.1.1): span selection on the GPU, an asynchronous pipeline and CUDA graphs for small batches: about 3 ms per text on an H100.
  • Same predictions: in fp32 the output matches the reference implementation exactly (agreement F1 1.000).

glinerel-fast against the reference implementation on an Apple M3 Pro: time per text on CPU and Apple GPU, one text and batches of 32, and loading time.

Pruning

Structured width pruning of SauerkrautLM-GLiNEREL-314m (narrower residual stream and feed-forward layers, factorized token embedding; all 14 layers and all attention heads kept), followed by knowledge distillation from SauerkrautLM-GLiNEREL-314m.

Architecture of the SauerkrautLM-GLiNEREL family: residual width, feed-forward width, token embedding, layers, heads, parameters and overall score of 314m, 128m and 72m.

Speed

glinerel-fast speed of the three SauerkrautLM-GLiNEREL sizes: CPU latency and throughput with and without torch.compile.

Time to process one German document of 1,065 characters, with the model warm and the median of 3 calls. On a laptop CPU, SauerkrautLM-GLiNEREL-72m takes 258 ms, gliner2.5-multi-v1 takes 1.3 s, and the 4B generative models take 13.5 and 14.5 minutes. Bars are linear up to 1.5 s, with an axis break for the generative models. One German document (1,065 characters), against gliner2.5 and 4B generative models.

Results

Zero-shot NER and relation F1 on the dev splits of CoNLL04, CrossRE, DocRED and FewRel. SauerkrautLM-GLiNEREL-72m has the best value in 10 of 11 rows; Qwen3.8-27B leads on CrossRE NER (77.7 vs 76.1). Its mean relation F1 (48.0) is 1.7× that of Qwen3.8-27B without reasoning. The LLM rows were scored by a separate scorer that follows the same protocol. Development splits, zero-shot, against gliner2.5, gliner-relex, GPT-6 Luna and Qwen3.8-27B.

Official test splits (CoNLL04, CrossRE) and multilingual NER and PII F1 of SauerkrautLM-GLiNEREL-314m, 128m and 72m.

Zero-shot strict relation F1 on the German RED^FM test split (285 documents, 32 relation types), with 95% document-level bootstrap intervals. All three SauerkrautLM-GLiNEREL sizes score 2.9 to 3.8 times gliner2.5-multi-v1, and the intervals do not overlap. Thresholds were chosen on the evaluated split, so scores are slightly optimistic. German relations (RED^FM test split).

Zero-shot strict entity F1 on the German test splits of RED^FM (285 documents) and MobIE (623 documents), with English and German label names. A SauerkrautLM-GLiNEREL model has the highest F1 in 9 of 10 rows, but on RED^FM all four models are within 3 points. Thresholds were chosen on the evaluated split. German entities (RED^FM and MobIE test splits).

Token usage, API latency and cost of the two LLM baselines on all 14,580 dev records, zero-shot with reasoning off. Costs are as reported by OpenRouter. Latency was measured with 16 calls in flight, so it is not comparable to local CPU time. Tokens, latency and cost of the LLM baselines.

LLM baselines with medium reasoning and four zero-shot instructions on a 600-record subset. Reasoning adds 2–3 overall points at 3 to 4.5 times the cost per record, and the best run (Qwen3.8-27B, p1) reaches 53.5. SauerkrautLM-GLiNEREL-72m's 61.8 was measured on the full dev splits, so the comparison is indicative. LLM baselines with reasoning and four prompts (600-record subset).

Thresholds are chosen per benchmark; LLMs get the same label sets with one fixed instruction at temperature 0.

Citation 

@misc{sauerkrautlm_glinerel_2026,
  title={SauerkrautLM-GLiNEREL: Joint Entity, PII and Relation Extraction in One Encoder Pass},
  author={Michele Montebovi and Hakim Bedhief},
  organization={VAGO Solutions},
  url={https://huggingface.co/VAGOsolutions/SauerkrautLM-GLiNEREL-72m},
  year={2026}
}

Family: SauerkrautLM-GLiNEREL-314m · SauerkrautLM-GLiNEREL-128m · SauerkrautLM-GLiNEREL-72m

Downloads last month
26
Safetensors
Model size
72.6M params
Tensor type
I64
·
F32
·
BF16
·
Inference Providers NEW
Token Classification
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for VAGOsolutions/SauerkrautLM-GLiNEREL-72m

Base model

LiquidAI/LFM2.5-230M-Base
Finetuned
LiquidAI/LFM2.5-Encoder-230M
Finetuned
VAGOsolutions/SauerkrautLM-GLiNEREL-314m
Finetuned
(2)
this model

Collection including VAGOsolutions/SauerkrautLM-GLiNEREL-72m