Turkish text. Different representations.
Two open models turn Turkish sentences into numbers that can be searched and compared by meaning. Choose a model for the work it will do.
- One sentence: Split into the pieces a model reads.
- BGE-M3 Distill 4L: Four Transformer layers; a summary token becomes 1024 numbers.
- TurboQuant 2-bit: No layers: each piece looks up 256 numbers stored in 2 bits; their mean is the sentence.
The sentence “Türkçe belgeler içinde arama.” splits into five pieces. BGE-M3 Distill 4L passes them through four Transformer layers, and its summary token becomes 1024 numbers (the model’s real output). Model2Vec TurboQuant 2-bit has no layers: each piece looks up its 256 numbers, stored in 2 bits, and their mean represents the sentence (illustrative values).
Evaluate Turkish retrieval models on your task.
Compact Turkish embedding models derived from BGE-M3 turn text into vectors for semantic search. Choosing between distilled and quantized variants involves evaluating retrieval quality on the target collection alongside model size and runtime requirements. Model cards explain the evaluation conditions behind published results.
Represent Turkish text as vectors
Encode Turkish sentences or passages as vectors for similarity comparison and semantic search.
Compare model options
Review model cards for architecture, size, intended use, and published evaluation conditions.
Measure task fit
Evaluate candidate embedding models on queries and documents that represent the target collection.
From input to output.
- Input: A Turkish search query and a collection of candidate passages.
- Process: Encode the query and passages with the selected embedding model, then rank them by vector similarity.
- Output: Ranked candidate passages to evaluate in a search or retrieval workflow.
In client projects.
ALTAI builds the model layer of search applications by comparing these models with alternatives on your organization's data. Selection is based on measured retrieval quality and the resource requirements of the target environment.
Choose a model for the work it will do.
BGE-M3 Distill 4L: 1024 Dimensions
- Architecture: Distilled Transformer · 4 layers
- Pooling: CLS
- Runtime: SentenceTransformers
- Evaluation: STS Pearson / Spearman on the card. Not a retrieval benchmark.
- License: Not on the card. Confirm with the publisher before use.
Model2Vec TurboQuant 2-bit: 256 Dimensions
- Architecture: Static embedding · 2-bit
- Pooling: Mean
- Runtime: TurboQuant · bundled in Akana
- Evaluation: STSb-TR and CPU throughput on the card. Measurement details are limited.
- License: Apache-2.0. Check the runtime and dependencies separately.
Start by choosing the runtime.
Pinned revision. Full inference was not tested here.
pip install sentence-transformers
python - <<'PY'
from sentence_transformers import SentenceTransformer
model = SentenceTransformer(
"altaidevorg/bge-m3-distill-4l",
revision="5309a3e2c4e8dd15514c2836d695afbff21797c6",
)
vectors = model.encode(["Türkçe belgeler içinde arama."])
print(vectors.shape) # expected: (1, 1024)
PY
Or the compact one.
Executed with the TurboQuant model bundled in Akana 0.6.0.
pip install akana==0.6.0
python - <<'PY'
import akana
# Akana’s bundled 2-bit TurboQuant model
vector = akana.embed("Türkçe belgeler içinde arama.")
print(len(vector)) # 256
print(akana.similarity(
"Filtreyi ne zaman temizlemeliyim?",
"Filtreyi her ay kontrol edin.",
))
PY
Files
- BGE-M3 Distill 4L: model.safetensors · config.json · modules.json
- TurboQuant 2-bit: turboquant_weights.npz · config.json
- Used in: letsearch, Akana
- Model cards: BGE-M3 distilled model card, 2-bit Turkish Model2Vec model card
- Demonstration: Training data and model demonstration