hs-embed-1
by hindsight-ai
A 112M-parameter multilingual text embedding model with 8,192-token inputs and 384-dimension vectors you can cut to 64
hindsight-ai/hs-embed-1Deploy hs-embed-1
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
hs-embed-1 is a text embedding model from Hindsight AI, published on Hugging Face on 28 October 2026 under the MIT license. It is built for the Hindsight agent memory project and works as a general bi-encoder: it maps text in English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese to 384-dimension vectors for semantic search, deduplication and retrieval for RAG. Queries and documents are encoded the same way, with no instruction prefix, and inputs can run to 8,192 tokens. Of its 112M parameters, 98M are the token embedding table, so each forward pass runs a 13M-parameter transformer, which makes it cheap to serve on CPU.
Architecture
A ModernBERT-family bi-encoder fine-tuned from hotchpotch/bekko-embedding-v1-a25m, with mean pooling and cosine similarity. It is trained with Matryoshka representation learning, so the first 256, 128 or 64 values of a vector work as a shorter embedding once re-normalized. Weights ship as FP16 safetensors (an FP32 copy is on the fp32 branch) and as an ONNX export with an int8 vocabulary table for CPU.
Mixpeek SDK Integration
# Embed documents with Sentence Transformers (usage from the model card).
# pip install -U sentence-transformers
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("hindsight-ai/hs-embed-1") # add truncate_dim=256 for shorter vectors
documents = [
"Scene 14: the presenter holds up the new phone and reads the price on screen.",
"Transcript, 00:03:12: we ship to Lisbon and Porto in two days.",
]
doc_vectors = model.encode(documents, normalize_embeddings=True)
# Store each vector with its file, timestamp and text in your vector index.Capabilities
- Multilingual text embeddings in 8 languages, including Simplified and Traditional Chinese
- Inputs up to 8,192 tokens, so long passages can be embedded without heavy chunking
- 384-dimension vectors that truncate to 256, 128 or 64 for smaller indexes
- The same encoding for queries and documents, with no prompt prefix to get wrong
- CPU serving through an int8 ONNX export, and GPU or CPU serving on Text Embeddings Inference
Use Cases on Mixpeek
Performance
We have not measured it, and the model card publishes its benchmark results only as charts, so no scores are quoted here. Score it on your own queries against your current model before switching.
Common Pipeline Companions
A larger multilingual model with dense, sparse and multi-vector outputs, for comparison on your own queries
Another small multilingual text embedder in the same size class
A widely used small multilingual baseline, which expects query and passage prefixes
Frequently Asked Questions
What is hs-embed-1?
A multilingual text embedding model from Hindsight AI, published on 28 October 2026 under the MIT license. It has 112M parameters, reads up to 8,192 tokens and returns 384-dimension vectors for semantic search and retrieval in English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese.
Does hs-embed-1 need a query prefix?
No. Queries and documents are encoded the same way, with no instruction or prefix. Normalize the vectors and compare them with cosine similarity.
Can I make hs-embed-1 vectors smaller?
Yes. It is trained with Matryoshka representation learning, so you can keep the first 256, 128 or 64 values of each 384-dimension vector and normalize again, or load it with truncate_dim in Sentence Transformers. Shorter vectors shrink the index; check recall on your own queries at each size.
Can hs-embed-1 search images or video?
Not directly, since it embeds text only. Index the text that comes out of your media, such as transcripts, captions, OCR and scene descriptions, and use a multimodal embedding model to search the pixels themselves.
Can I run hs-embed-1 on a CPU?
Yes. Its forward pass runs a 13M-parameter transformer, and it ships an int8 ONNX export that Text Embeddings Inference uses on its CPU image.
Specification
Build a pipeline with hs-embed-1
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free