Model2Vec.Net 0.1.1

Model2Vec.Net

A pure-managed C# (net10.0) port of MinishLab Model2Vec static-embedding inference. Load a Model2Vec model folder (model.safetensors + tokenizer.json + config.json) and compute sentence embeddings with no Python, no native libraries, and no ONNX.

Unofficial, independent port. Not affiliated with or endorsed by MinishLab. See the project repository for full attribution and third-party notices.

Features

• Pure C# (net10.0), no native dependency, no P/Invoke.
• Reads the safetensors format directly (F32, F16, F64, I8, U8).
• Supports Model2Vec embeddings and Sentence Transformers embedding.weight tensors, plus Model2Vec vocabulary quantization (weights/mapping).
• Hugging Face tokenizers via Microsoft.ML.Tokenizers: WordPiece, byte-level BPE, and SentencePiece-backed Unigram.
• SIMD-accelerated scaling and normalization via System.Numerics.Tensors.
• Output verified against the Python model2vec package (tolerance 1e-4).

Installation

dotnet add package Model2Vec.Net

Usage

using Model2VecNet;

var model = Model2VecModel.Load(@"C:\models\potion-base-2M");

float[] embedding = model.Encode("The quick brown fox jumps over the lazy dog.");

Console.WriteLine(model.Dimension);   // e.g. 256
Console.WriteLine(embedding.Length);  // == Dimension

Batch encoding:

float[][] embeddings = model.Encode([
    "First sentence",
    "Second sentence",
]);

Model2VecModel is immutable after loading and safe to share across threads.

Microsoft.Extensions.AI

Model2VecModel implements Microsoft.Extensions.AI IEmbeddingGenerator<string, Embedding<float>>, so it plugs directly into the .NET AI ecosystem (RAG pipelines, vector stores, semantic search):

using Microsoft.Extensions.AI;
using Model2VecNet;

IEmbeddingGenerator<string, Embedding<float>> generator = Model2VecModel.Load(@"C:\models\potion-base-2M");

GeneratedEmbeddings<Embedding<float>> embeddings = await generator.GenerateAsync(["First sentence", "Second sentence"]);
ReadOnlyMemory<float> vector = embeddings[0].Vector;

This composes with any Microsoft.Extensions.VectorData store (for example Hnsw.Net) to embed and index text with no external service.

Scope: inference only

This package loads a distilled Model2Vec static model and encodes text. It does not distill or train models — that is, it does not forward-pass a teacher sentence-transformer over a vocabulary, run PCA, apply Zipf/SIF weighting, or run the tokenlearn/classifier training steps.

Why: those steps require a full transformer encoder plus an autodiff/optimizer training loop, which in .NET means a native deep-learning dependency (ONNX Runtime or libtorch). That would break this package's pure-managed, no-native-dependency design. Distillation is a one-time offline step — create the model once with the upstream Python tooling and load the resulting model.safetensors here.

Getting a model

Models are published on Hugging Face and are not bundled with this package. Download a model folder (containing model.safetensors, tokenizer.json, and config.json) and pass its path to Model2VecModel.Load. For example:

• minishlab/potion-base-2M
• minishlab/potion-base-8M

Any Model2Vec model in the standard folder layout will load.

Upstream / attribution

• Model2Vec (reference implementation): https://github.com/MinishLab/model2vec
• Microsoft.ML.Tokenizers: https://github.com/dotnet/machinelearning-tokenizers

License

MIT. See the project repository for the license and third-party notices: https://github.com/ericstj/Model2Vec.Net