🚀 vtx-embed-7M (mini)

The world's most memory-efficient static embedding model powering vortexa. Native 4-Bit quantization · 4.72 MB RAM · 7.56M Parameters · Matryoshka MRL · Sub-millisecond CPU latency

---

⚡ Integration with Vortexa

This model powers vortexa — a standalone codebase indexing and semantic search engine designed for AI agents and developers.

vortexa builds a persistent, hybrid search index over source code using:

• Dense Retrieval: Driven natively by vtx-embed-7M (on-the-fly LF4 4-bit dequantization, SIF+PC pooling, Matryoshka truncation).
• Sparse Retrieval: BM25 keyword scoring for exact symbol matches.
• AST-Aware Chunking: Tree-sitter powered chunking respecting function and class boundaries.
• LMDB Storage: Fast, persistent vector and document chunk storage.

---

📄 Model Details

---

📊 Official Benchmark Results

---

💻 Quickstart Usage

Native Vortexa Core API

from vortexa.core.inference import VortexEmbedInference, similarity

# Load model using the "mini" model alias
model = VortexEmbedInference("mini")

queries = [
    "What is the capital of India?",
    "Explain gravity and general relativity",
]
documents = [
    "The capital of India is New Delhi.",
    "Gravity is a fundamental interaction that causes mutual attraction between all things with mass or energy.",
]

# 1. Encode queries and documents
query_embeddings = model.encode(queries)
document_embeddings = model.encode(documents)

# 2. Compute similarity matrix directly
similarity_matrix = query_embeddings @ document_embeddings.T
print("Similarity Matrix:")
print(similarity_matrix)
# Example output:
# [[0.82, 0.12],
#  [0.11, 0.74]]

# 3. Use built-in model similarity method
scores = model.similarity(query_embeddings, document_embeddings)

# 4. Single query against document list lookup
scores_single = model.similarity("What is the capital of India?", documents)
print("Single query scores:", scores_single)

---

📜 Citation

@misc{vtx-embed-7m},
  title  = {vtx-embed-7M: Native 4-Bit Embeddings for Standalone Codebase Indexing},
  author = {VTXAI},
  year   = {2026},
  url    = {https://huggingface.co/VTXAI/vtx-embed-7M}
}

---

📄 License

MIT License — free for commercial and research use.