hs-embed-1

hs-embed-1 is a multilingual text embedding model built for Hindsight.

Highlights

• Multilingual: English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese (Simplified and Traditional).
• Variable dimensions: 384-dimensional vectors that can be truncated to 256, 128 or 64.
• Long inputs: up to 8,192 tokens.
• Small: 112M parameters. 98M of them are the token embedding table, so each forward pass runs a 13M-parameter transformer.
• No prompts: queries and documents are encoded the same way, with no prefix.
• CPU and GPU: ships an int8 ONNX export for CPU, and safetensors for GPU.
• Serves on TEI: CPU and GPU images.

Model overview

Model type	Text embedding (bi-encoder)
Base model	hotchpotch/bekko-embedding-v1-a25m
Parameters	112M
Languages	en, es, fr, de, pt, ja, ko, zh
Context length	8,192 tokens
Embedding dimensions	384 (truncatable to 256, 128, 64)
Pooling	Mean
Similarity	Cosine
Formats	safetensors stored in FP16 (FP32 copy on the fp32 branch), ONNX (onnx/model.onnx, int8 vocabulary table)

Evaluation

Serving on CPU

Usage

Sentence Transformers

pip install -U sentence-transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("hindsight-ai/hs-embed-1")

queries = ["Where did Maya move last spring?"]
documents = [
    "Maya moved to Lisbon in April.",
    "Maya's brother lives in Porto.",
]

query_embeddings = model.encode(queries, normalize_embeddings=True)
document_embeddings = model.encode(documents, normalize_embeddings=True)

print(model.similarity(query_embeddings, document_embeddings))

For shorter vectors, load the model with truncate_dim:

model = SentenceTransformer("hindsight-ai/hs-embed-1", truncate_dim=256)

Text Embeddings Inference

CPU (uses onnx/model.onnx):

docker run -p 8080:80 \
  ghcr.io/huggingface/text-embeddings-inference:cpu-1.9.3 \
  --model-id hindsight-ai/hs-embed-1 --auto-truncate

GPU:

docker run --gpus all -p 8080:80 \
  ghcr.io/huggingface/text-embeddings-inference:89-1.9.3 \
  --model-id hindsight-ai/hs-embed-1 --auto-truncate

Use the GPU image tag for your GPU: 89-1.9.3 for L4 and other Ada GPUs, 1.9.3 for A100 and H100.

curl 127.0.0.1:8080/embed \
  -H 'Content-Type: application/json' \
  -d '{"inputs": ["Where did Maya move last spring?", "Maya moved to Lisbon in April."]}'

To use fewer dimensions, keep the first 256, 128 or 64 values of each vector and normalize again.

Hindsight

HINDSIGHT_API_EMBEDDINGS_PROVIDER=local
HINDSIGHT_API_EMBEDDINGS_LOCAL_MODEL=hindsight-ai/hs-embed-1

Or point Hindsight at a TEI server:

HINDSIGHT_API_EMBEDDINGS_PROVIDER=tei
HINDSIGHT_API_EMBEDDINGS_TEI_URL=http://localhost:8080

Hindsight Inference

hs-embed-1 is also served at https://inference.hindsight.vectorize.io. To use it from Hindsight:

HINDSIGHT_API_EMBEDDINGS_PROVIDER=hindsight
HINDSIGHT_API_EMBEDDINGS_HINDSIGHT_API_KEY=<your Hindsight Inference API key>

License

MIT

Citation

@misc{hindsightai2026hsembed1,
  title  = {hs-embed-1},
  author = {{Hindsight AI}},
  year   = {2026},
  url    = {https://huggingface.co/hindsight-ai/hs-embed-1}
}