Knowledge Distillation from strong synthetic data (200k samples generated with Qwen/Qwen3-32B-AWQ and scored by a high-performing reranker). Robust 210M parameter architecture optimized for multilingual reasoning-focused retrieval without compression trade-offs.

🎯 Core Features and Innovations:

Next-Generation Knowledge Distillation: By utilizing 200,000 synthetically generated, high-quality training examples (created with Qwen/Qwen3-32B-AWQ and scored by a state-of-the-art reranker), our model learns complex reasoning patterns from models 54× its size.
Optimized Architecture: Full 210M parameters preserve maximum capacity for complex reasoning patterns

💪 David vs. Goliath: Small but Mighty

With 210 million parameters – that's less than 1/33rd the size of some competing models – SauerkrautLM-Reason-EuroColBERT achieves or exceeds the performance of:

Models with over 7 billion parameters (33× larger than ours)
Proprietary API-based solutions from major tech companies
Specialized reasoning models like ReasonIR-8B (38× larger)

This balanced architecture provides exceptional performance while remaining deployable on standard infrastructure.

Model Overview

Model: VAGOsolutions/SauerkrautLM-Reason-EuroColBERT
Base: Fine-tuned from VAGOsolutions/SauerkrautLM-EuroColBERT using knowledge distillation
Architecture: PyLate / ColBERT (Late Interaction)
Languages: Multilingual (optimized for 7 European languages: German, English, Spanish, French, Italian, Dutch, Portuguese)
License: Apache 2.0
Model Size: 210M parameters Efficiency Ratio: Up to 38× smaller than comparable performing models

Model Description

Model Type: Multi-vector embedding model with innovative Late Interaction architecture
Document Length: 8192 tokens (32× longer than traditional BERT models)
Query Length: 256 tokens (optimized for complex, multi-part queries)
Output Dimensionality: 128 tokens (efficient vector representation)
Similarity Function: MaxSim (enables precise token-level matching)
Training Loss: Knowledge Distillation (PyLate)

Architecture

ColBERT(
  (0): Transformer(ModernBertModel)
  (1): Dense(768 -> 128 dim, no bias)
)

🔬 Technical Innovations in Detail

Knowledge Distillation: The Student Surpassing the Master

Our 210M parameter model leverages state-of-the-art knowledge distillation:

Synthetic Data Generation: 200,000 high-quality query-document pairs generated using the Qwen/Qwen3-32B-AWQ model (32 billion parameters) based on the ReasonIR approach
Quality Assurance: Each pair evaluated and filtered by a state-of-the-art reranker
Distillation Process: The EuroColBERT model learns to replicate the ranking patterns of large models while maintaining its multilingual strengths

Architectural Advantages

SauerkrautLM-Reason-EuroColBERT leverages its full 210M parameters to deliver:

Superior multilingual performance: Native optimization for 7 European languages
No compression trade-offs: Full parameter capacity ensures maximum reasoning capability
Balanced efficiency: 33-38× smaller than large models while maintaining competitive performance

This architecture combines the advantages of Late Interaction Retrieval (precise token-level matching) with robust multilingual capabilities.

🔬 Benchmarks: David vs. Goliath Performance

Our comprehensive evaluation demonstrates that model size is not destiny. Despite being 33-38× smaller than competing models, SauerkrautLM-Reason-EuroColBERT consistently delivers superior or comparable performance across challenging reasoning and multilingual retrieval tasks.

BRIGHT Benchmark (English, reasoning‑focused retrieval)

The BRIGHT benchmark is designed to evaluate reasoning‑intensive retrieval. All scores are nDCG@10. SauerkrautLM-Reason-EuroColBERT (210M parameters) is compared with dense and proprietary baselines as well as other SauerkrautLM variants.

Model / Metric	Biology	Earth	Economics	Psychology	Robotics	Stackoverflow	Sustainable	Leetcode	Pony	AoPS	Theorem‑Q	Theorem‑T	Mean StackEx	Mean coding	Mean theorem	Full Mean
BM25	18.90	27.20	14.90	12.50	13.60	18.40	15.00	24.40	7.90	6.20	10.40	4.90	17.21	16.15	7.17	14.53
< 1 B OS
BGE	11.70	24.60	16.60	17.50	11.70	10.80	13.30	26.70	5.70	6.00	13.00	6.90	15.17	16.20	8.63	13.71
Inst‑L	15.20	21.20	14.70	22.30	11.40	13.30	13.50	19.50	1.30	8.10	20.90	9.10	15.94	10.40	12.70	14.21
SBERT	15.10	20.40	16.60	22.70	8.20	11.00	15.30	26.40	7.00	5.30	20.00	10.80	15.61	16.70	12.03	14.90
> 1 B OS
E5	18.60	26.00	15.50	15.80	16.30	11.20	18.10	28.70	4.90	7.10	26.10	26.80	17.36	16.80	20.00	17.93
SFR	19.10	26.70	17.80	19.00	16.30	14.40	19.20	27.40	2.00	7.40	24.30	26.00	18.93	14.70	19.23	18.30
Inst‑XL	21.60	34.30	22.40	27.40	18.20	21.20	19.10	27.50	5.00	8.50	15.60	5.90	23.46	16.25	10.00	18.89
GritLM	24.80	32.30	18.90	19.80	17.10	13.60	17.80	29.90	22.00	8.80	25.20	21.20	20.61	25.95	18.40	20.95
Qwen	30.60	36.40	17.80	24.60	13.20	22.20	14.80	25.50	9.90	14.40	27.80	32.90	22.80	17.70	25.03	22.51
Proprietary
Cohere	18.70	28.40	20.40	21.60	16.30	18.30	17.60	26.80	1.90	6.30	15.70	7.20	20.19	14.35	9.73	16.60
OpenAI	23.30	26.70	19.50	27.60	12.80	14.30	20.50	23.60	2.40	8.50	23.50	11.70	20.67	13.00	14.57	17.87
Voyage	23.10	25.40	19.90	24.90	10.80	16.80	15.40	30.60	1.50	7.50	27.40	11.60	19.47	16.05	15.50	17.91
Google	22.70	34.80	19.60	27.80	15.70	20.10	17.10	29.60	3.60	9.30	23.80	15.90	22.54	16.60	16.33	20.00
ReasonIR data
ReasonIR‑8B	26.20	31.40	23.30	30.00	18.00	23.90	20.50	35.00	10.50	14.70	31.90	27.20	24.76	22.75	24.60	24.38
Reason‑ModernColBERT (149 M) reported	33.25	41.02	24.93	30.73	21.12	20.62	20.31	31.07	8.51	9.17	19.51	11.24	27.43	19.79	15.38	22.62
Reason‑ModernColBERT (149 M) our eval**	34.28	41.53	19.96	27.02	21.15	23.62	17.21	26.61	1.32	7.30	19.79	9.70	27.93	13.97	12.26	20.79
SauerkrautLM Reasoning data
SauerkrautLM-Multi-Reason-ModernColBERT (149 M)	36.92	45.53	19.47	27.04	19.35	25.31	20.78	29.74	12.54	10.52	14.62	7.65	28.94	21.14	10.93	22.45
SauerkrautLM‑Reason‑EuroColBERT (210 M)	38.16	39.43	16.99	24.49	17.50	17.60	20.72	29.10	13.57	12.04	10.43	4.95	25.70	21.33	9.14	20.42
SauerkrautLM‑Reason‑Multi‑ColBERT (15 M)	23.33	23.78	10.53	9.03	10.28	10.88	13.13	18.10	15.86	1.75	4.29	0.81	14.64	16.98	2.28	11.81

Evaluation note: our re‑evaluation of Reason‑ModernColBERT uses the same query‑length settings from the original Lighton repo; the instructions for the originally reported scores are not public.

⚖️ Relative Efficiency

With 210M parameters, SauerkrautLM-Reason-EuroColBERT demonstrates that balanced architecture design can surpass several ≥7B dense and proprietary retrievers on reasoning‑centric tasks while maintaining excellent multilingual performance.

BRIGHT Benchmark (German, reasoning‑focused retrieval)

All scores are nDCG@10.

Model / Metric	Biology	Earth	Economics	Psychology	Robotics	Stackoverflow	Sustainable	Leetcode	Pony	AoPS	Theorem‑Q	Theorem‑T	Mean StackEx	Mean coding	Mean theorem	Full Mean
SauerkrautLM‑Multi‑Reason‑ModernColBERT (149 M)	28.00	34.71	12.90	17.98	13.67	19.64	17.70	11.66	15.49	7.27	6.76	1.32	21.15	13.57	5.11	15.59
SauerkrautLM‑Reason‑EuroColBERT (210 M)	31.09	31.48	11.95	18.39	11.25	14.43	20.26	25.67	12.15	9.58	8.15	2.76	19.76	18.91	6.83	16.43
SauerkrautLM‑Reason‑Multi‑ColBERT (15 M)	15.37	20.11	7.36	7.07	4.24	4.71	7.67	0.77	6.31	3.81	0.76	0.00	9.81	3.54	1.52	6.51

Observation: The 210M EuroColBERT model secures the highest Full‑Mean (16.43) across German benchmarks, with particularly strong performance on coding (18.91) and theorem proving (6.83) tasks, demonstrating its superior multilingual reasoning capabilities.

NanoBEIR Europe (multilingual retrieval)

Average nDCG@10 across the seven languages we evaluated:

Language	nDCG@10
de	47.71
en	58.72
es	52.15
fr	50.46
it	49.85
nl	48.47
pt	50.72

Why SauerkrautLM Matters for Production

Outperforms proprietary APIs: beats Cohere, OpenAI, Voyage and Google on BRIGHT Full Mean while remaining fully open‑source under a permissive Apache 2.0 license.
Superior multilingual performance with the highest German BRIGHT Full-Mean (16.43) — demonstrating exceptional cross-lingual reasoning capabilities.
Full parameter range: from the tiny 15 M Multi‑ColBERT (competitive with SBERT‑scale encoders) to the robust 210 M EuroColBERT variant.
Matches or exceeds models 33–38× larger (e.g. ReasonIR‑8B, GritLM-7B, Qwen-7B).
Strong multilingual coverage across seven European languages without language‑specific fine‑tuning.

We translated both BRIGHT and NanoBEIR into seven European languages to rigorously evaluate multilingual retrieval capabilities.

Below is a scatter plot that visualises model size (millions of parameters) against BRIGHT Full‑Mean nDCG@10. SauerkrautLM models occupy the best trade‑off region—smallest models with top‑tier reasoning performance.

Real-World Impact

The efficiency gains translate to tangible benefits:

Democratized AI: Run state-of-the-art retrieval on consumer hardware
Edge Deployment: Enable on-device search for privacy-sensitive applications
Massive Scale: Index billions of documents at a fraction of traditional costs

📈 Summary: Balanced Excellence in Multilingual Retrieval

SauerkrautLM-Reason-EuroColBERT represents the optimal balance between model size and performance. By combining cutting-edge knowledge distillation with a robust 210M parameter architecture, we've created a model that:

Achieves the highest German BRIGHT Full-Mean (16.43) among all SauerkrautLM variants
Excels at multilingual reasoning with particularly strong performance on coding and theorem-proving tasks
Outperforms models 33-38× larger while maintaining manageable infrastructure requirements
Delivers superior multilingual coverage across 7 European languages
Provides production-ready performance without the extreme compression trade-offs

This model demonstrates that the EuroBERT architecture design at 210M parameters can deliver exceptional multilingual reasoning capabilities while remaining practical for real-world deployment.

Model

This is a multi-vector (ColBERT-style late interaction) embedding model. It maps sentences & paragraphs to sequences of 128-dimensional dense vectors and can be used for semantic textual similarity using the MaxSim operator.

Usage

Sentence Transformers

This model can be used with Sentence Transformers as a multi-vector (ColBERT-style late interaction) retriever via the MultiVectorEncoder:

pip install "sentence-transformers>=6.0.0"

python

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("VAGOsolutions/SauerkrautLM-Reason-EuroColBERT")

query = "Which planet is known as the Red Planet?"
documents = [
    "Venus is often called Earth's twin because of its similar size and proximity.",
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Jupiter is the largest planet in our solar system.",
    "Saturn is famous for its beautiful rings.",
]

query_embeddings = model.encode_query(query)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings[0].shape)
# (256, 128) (17, 128)

# MaxSim late-interaction scoring (higher is more relevant)
scores = model.similarity(query_embeddings, document_embeddings)
print(scores)
# tensor([[244.6474, 244.7671, 244.5400, 241.9365]])

PyLate

First install the PyLate library:

pip install -U pylate

Retrieval

PyLate provides a streamlined interface to index and retrieve documents using ColBERT models. The index leverages the Voyager HNSW index to efficiently handle document embeddings and enable fast retrieval.

Indexing documents

First, load the ColBERT model and initialize the Voyager index, then encode and index your documents:

python

from pylate import indexes, models, retrieve

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path="VAGOsolutions/SauerkrautLM-Reason-EuroColBERT",
)

# Step 2: Initialize the Voyager index
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
    override=True,  # This overwrites the existing index if any
)

# Step 3: Encode the documents
documents_ids = ["1", "2", "3"]
documents = ["document 1 text", "document 2 text", "document 3 text"]

documents_embeddings = model.encode(
    documents,
    batch_size=32,
    is_query=False,  # Ensure that it is set to False to indicate that these are documents, not queries
    show_progress_bar=True,
)

# Step 4: Add document embeddings to the index by providing embeddings and corresponding ids
index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

Note that you do not have to recreate the index and encode the documents every time. Once you have created an index and added the documents, you can re-use the index later by loading it:

python

# To load an index, simply instantiate it with the correct folder/name and without overriding it
index = indexes.Voyager(
    index_folder="pylate-index",
    index_name="index",
)

Retrieving top-k documents for queries

Once the documents are indexed, you can retrieve the top-k most relevant documents for a given set of queries. To do so, initialize the ColBERT retriever with the index you want to search in, encode the queries and then retrieve the top-k documents to get the top matches ids and relevance scores:

python

# Step 1: Initialize the ColBERT retriever
retriever = retrieve.ColBERT(index=index)

# Step 2: Encode the queries
queries_embeddings = model.encode(
    ["query for document 3", "query for document 1"],
    batch_size=32,
    is_query=True,  #  # Ensure that it is set to False to indicate that these are queries
    show_progress_bar=True,
)

# Step 3: Retrieve top-k documents
scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=10,  # Retrieve the top 10 matches for each query
)

Reranking

If you only want to use the ColBERT model to perform reranking on top of your first-stage retrieval pipeline without building an index, you can simply use rank function and pass the queries and documents to rerank:

python

from pylate import rank, models

queries = [
    "query A",
    "query B",
]

documents = [
    ["document A", "document B"],
    ["document 1", "document C", "document B"],
]

documents_ids = [
    [1, 2],
    [1, 3, 2],
]

model = models.ColBERT(
    model_name_or_path="VAGOsolutions/SauerkrautLM-Reason-EuroColBERT",
)

queries_embeddings = model.encode(
    queries,
    is_query=True,
)

documents_embeddings = model.encode(
    documents,
    is_query=False,
)

reranked_documents = rank.rerank(
    documents_ids=documents_ids,
    queries_embeddings=queries_embeddings,
    documents_embeddings=documents_embeddings,
)

Citation

BibTeX

SauerkrautLM‑Reason‑EuroColBERT

bibtex

@misc{SauerkrautLM-Reason-EuroColBERT,
  title={SauerkrautLM-Reason-EuroColBERT},
  author={David Golchinfar},
  url={https://huggingface.co/VAGOsolutions/SauerkrautLM-Reason-EuroColBERT},
  year={2025}
}

EuroBERT-210m

bibtex

@misc{boizard2025eurobertscalingmultilingualencoders,
      title={EuroBERT: Scaling Multilingual Encoders for European Languages}, 
      author={Nicolas Boizard and Hippolyte Gisserot-Boukhlef and Duarte M. Alves and André Martins and Ayoub Hammal and Caio Corro and Céline Hudelot and Emmanuel Malherbe and Etienne Malaboeuf and Fanny Jourdan and Gabriel Hautreux and João Alves and Kevin El-Haddad and Manuel Faysse and Maxime Peyrard and Nuno M. Guerreiro and Patrick Fernandes and Ricardo Rei and Pierre Colombo},
      year={2025},
      eprint={2503.05500},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.05500}, 
}
}

Sentence Transformers

bibtex

@inproceedings{reimers-2019-sentence-bert,
  title = {Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks},
  author = {Reimers, Nils and Gurevych, Iryna},
  booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing},
  month = {11},
  year = {2019},
  publisher = {Association for Computational Linguistics},
  url = {https://arxiv.org/abs/1908.10084}
}

PyLate

bibtex

@misc{PyLate,
  title={PyLate: Flexible Training and Retrieval for Late Interaction Models},
  author={Chaffin, Antoine and Sourty, Raphaël},
  url={https://github.com/lightonai/pylate},
  year={2024}
}

Acknowledgements

We thank Antoine Chaffin (LightOn AI) for helpful discussions and for clarifying evaluation settings for Reason‑ModernColBERT, and the PyLate team for providing the training framework that made this work possible.