A 6-layer MiniLM sentence encoder that maps text to a 384-dimensional normalized vector. It is the default workhorse for semantic search, clustering, deduplication and RAG retrieval — roughly 90 MB and fast enough to embed thousands of sentences per second on CPU.

[!NOTE] This is a mirror. The weights and tokenizer files here are an unmodified copy of sentence-transformers/all-MiniLM-L6-v2, re-hosted on this profile for reproducibility and convenience. All credit for the original work belongs to its authors. The upstream license (apache-2.0) is preserved and applies to this copy. If you need the canonical version, please use the upstream repository.

This mirror carries only the PyTorch/safetensors weights and tokenizer (~92 MB). The upstream ONNX, OpenVINO, TensorFlow and Rust variants were intentionally left out to keep the repo small; fetch those from upstream if you need them.

Specs


Layers	6
Embedding dim	384
Max sequence length	256 word pieces (longer input is truncated)
Pooling	mean, then L2-normalized
Parameters	~22.7M

Usage

With sentence-transformers (recommended — pooling and normalization are handled for you):

python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("priyaganesh2050/all-MiniLM-L6-v2")
emb = model.encode(["How do I reset my password?", "password recovery steps"])
print(emb.shape)                       # (2, 384)
print(model.similarity(emb[0], emb[1]))

With plain transformers:

python

import torch, torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel

tok = AutoTokenizer.from_pretrained("priyaganesh2050/all-MiniLM-L6-v2")
model = AutoModel.from_pretrained("priyaganesh2050/all-MiniLM-L6-v2")

def embed(texts):
    batch = tok(texts, padding=True, truncation=True, return_tensors="pt")
    out = model(**batch).last_hidden_state
    mask = batch["attention_mask"].unsqueeze(-1).float()
    pooled = (out * mask).sum(1) / mask.sum(1).clamp(min=1e-9)
    return F.normalize(pooled, p=2, dim=1)

print(embed(["semantic search", "vector retrieval"]) @ embed(["finding similar text"]).T)

Notes

Embeddings are already unit length, so cosine similarity is just a dot product.
Input beyond 256 word pieces is silently truncated — chunk long documents before embedding.
English-only. For multilingual work use a multilingual encoder instead.