Downloads · 30 days
22
14% of all-time downloads
andreasmartin/apertus-v1.1-swiss-embed-0.4b-bidir-GGUF
apertus-v1.1-swiss-embed-0.4b-bidir-GGUF is a feature extraction model from andreasmartin. Use it when you need embeddings to search or compare text. It is set up for gguf. The card lists the license as apache-2.0.
Stock-llama.cpp-compatible GGUF conversion of andreasmartin/apertus-v1.1-swiss-embed-0.4b-bidir.
Downloads · 30 days
22
14% of all-time downloads
All-time downloads
152
Public
Repo size
625 MB
Likes
0
Public
Click a slice to open those files.
.gguf616 MB · 99%
From the Hugging Face model README
Stock-llama.cpp-compatible GGUF conversion of andreasmartin/apertus-v1.1-swiss-embed-0.4b-bidir.
This is a bidirectional Apertus sentence-embedding model, not a causal language model.
Original Sentence Transformers pipeline:
ApertusModel → mean pooling → Dense projection → L2 normalization
Embedding dimension: 1024.
No Modelfile is required. The GGUF can be loaded directly from Hugging Face:
ollama run hf.co/andreasmartin/apertus-v1.1-swiss-embed-0.4b-bidir-GGUF:Q8_0
Release GGUF: apertus-v1.1-swiss-embed-0.4b-bidir-Q8_0.gguf
The stock Apertus runtime can execute the bidirectional Apertus backbone and mean pooling,
but it does not currently bind the Sentence Transformers dense_2.weight tensor.
Therefore the GGUF intentionally contains the stock-compatible pre-Dense embedding. To reproduce the original Sentence Transformers embedding, apply:
y = L2_normalize(x @ W.T)
where:
x is the 1024-dimensional embedding returned by the stock runtimeW is dense_2.weightThe repository includes:
dense_2.weight.f32.bin — raw float32 matrixdense_2.weight.npy — NumPy representationdense-postprocess.json — dimensions and operation metadataThe stock GGUF was compared with the full reference pipeline on 12 multilingual and semantically diverse texts after applying the original Dense projection.
Validation verdict:
STOCK GGUF + DENSE POSTPROCESS REPRODUCES REFERENCE
apertus-v1.1-swiss-embed-0.4b-bidir-Q8_0.gguf4d9176092d00586775af140581bb0b558ddc4389attention.causal = falseoutput.weight included for stock Apertus loader compatibility