hs-embed-1

hs-embed-1

Turns text into compact vectors for search and matching in eight languages, with inputs up to 8,192 tokens. A small 112M embedding model built for Hindsight, running on CPU or GPU.

Made by hindsight-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: hs-embed-1 turns text in eight languages into vectors for search, and hindsight-ai made it for use with Hindsight. Each sentence or passage becomes a 384-number vector, and you can shorten that vector to 256, 128 or 64 numbers. It has 112M parameters, accepts inputs up to 8,192 tokens, and needs no prompt prefix for queries or documents.

Who it suits: Developers who build semantic search, retrieval or agent memory and want a small embedding model that runs on a CPU. The files carry the MIT license.

Industry | Who uses it | How they use it | Example
Search | Search engineers | Embed queries and documents for semantic search in eight languages | Matching a question about a move to a sentence naming a city
AI agents | Agent developers | Give an agent memory by pointing Hindsight at this model | Setting Hindsight to use the local model
Platform | Infrastructure teams | Serve embeddings from a Text Embeddings Inference server | Running the CPU image with Docker

Anyone who needs a huge embedding model with the highest scores can skip it, since this one is built to be small.

Where to use it:

• Python with the sentence-transformers library
• A Text Embeddings Inference server, on CPU or GPU
• Hindsight, with a local model, a TEI server or the hosted Hindsight Inference service
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Semantic search | Developers | Vectors you compare with cosine similarity
Agent memory | Agent builders | Embeddings that Hindsight stores and searches
Short vectors | Cost-conscious teams | Vectors cut to 256, 128 or 64 numbers

How to set it up:

• Python and the sentence-transformers package
• Docker, only if you serve the model with Text Embeddings Inference
• A Hindsight API key, only for the hosted Hindsight Inference service

1. Install Python.
2. Install the sentence-transformers package with the first block below.
3. Optionally copy the model files with Hugging Face's two commands in the second block.

pip install -U sentence-transformers

pip install -U huggingface_hub
hf download hindsight-ai/hs-embed-1

How to use it: You load the model, encode your queries and documents with normalized embeddings, and compare them with the similarity method.

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))

Passing truncate_dim when you load the model gives shorter vectors.

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

To make Hindsight use the model on your machine, set these two environment variables.

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

Tool | What it does
SentenceTransformer | Loads the model and encodes text
truncate_dim | Shortens the vectors to 256, 128 or 64 numbers
Text Embeddings Inference | Serves the model over HTTP on CPU or GPU
Hindsight settings | Tell Hindsight which embeddings provider to use

Example: You ask which city Maya moved to, using the question "Where did Maya move last spring?" against two sentences, one about Lisbon and one about her brother in Porto. The call prints a similarity score for each sentence, and the Lisbon one should score higher.

For newcomers and experts: An embedding turns text into numbers so that similar meanings sit close together. A good first step is to run the Python example above and read the two scores it prints.

Most of the 112M parameters, 98M, sit in the token embedding table, which leaves a transformer of just 13M parameters for each pass. Pooling is a mean, similarity is cosine, and the model ships an int8 ONNX export for CPU alongside safetensors for GPU. To shorten vectors by hand, keep the first values and normalize again.

Load hs-embed-1
Encode queries and documents
Normalize the vectors
Compare them with cosine similarity
Rank the matches

Good to know:

• License: MIT
• Cost to run: none named for local use, while the hosted Hindsight Inference service needs an API key
• Data it reaches: the download fetches files from Hugging Face, and the hosted service receives whatever text you send it
• Eight languages, among them English, Japanese and Chinese in both scripts

Title: Vector sizes you can use
Full vector: 384 dims
Truncated: 256 dims
Truncated again: 128 dims
Smallest: 64 dims

On Hugging Face, 13 November 2026

• Task: sentence similarity
• Library: sentence-transformers
• Parameters: 112M
• Weights: about 223 MB in F16; about 56 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 22,412
• Likes: 7
• License: mit
• Languages: en, es, fr, de, pt, ja, ko, zh
• Based on: hotchpotch/bekko-embedding-v1-a25m
• Tags: sentence-transformers, onnx, safetensors, modernbert, feature-extraction, embeddings, matryoshka, multilingual, text-embeddings-inference, hindsight, sentence-similarity

Links

• Model: https://huggingface.co/hindsight-ai/hs-embed-1
• Files: https://huggingface.co/hindsight-ai/hs-embed-1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hindsight-ai/hs-embed-1

Then follow the usage in its model card. Nothing in this kit runs by itself.