Minish
Minish is a two-person open-source lab building small, efficient, effective natural-language-processing models and packages. Its stated goal is to make state-of-the-art NLP accessible regardless of hardware budget or deployment constraints.
At a Glance
- NLP and machine-learning engineers
- Developers building search, RAG, classification, and retrieval systems
- Teams curating or training large language-model datasets
- Companies needing CPU, edge, mobile, or low-cost inference
- +2 more
AI Tools by Minish
(1)
Semble
Open Source Code Search for Agents
Discussions
No discussions yet
Be the first to start a discussion about Minish
Latest News
Model2Vec v0.10.0 adds training features, ONNX export, and bug fixes
Semble launches as a fast, CPU-only code-search library for AI agents
Model2Vec size improvements introduce vocabulary quantization and demonstrate models as small as about 6 MB
Minish compares Model2Vec with fastText, reporting faster inference, smaller models, and stronger average classification results
Products & Services
A technique and Python package that distills sentence transformers into compact static embedding models. It can reduce model size by up to 50x and make inference up to 500x faster, using fixed token vectors, PCA, weighting, and optional quantization.
Open-source pretrained Model2Vec static embedding models, including potion-base-2M, potion-base-4M, and potion-base-8M, trained with the Tokenlearn approach.
A lightweight multimodal library for semantic deduplication, outlier filtering, and representative sample selection across text, images, audio, and custom encoders.
A lightweight, low-dependency vector store with a unified interface for nearest-neighbor backends, distance metrics, indexing, and evaluation.
Market Position
Minish positions its products as small, fast, and resource-efficient alternatives to transformer embedding stacks. Model2Vec is compared with sentence transformers, fastText, GloVe, and BPEmb; SemHash addresses semantic similarity that character- or word-n-gram methods such as MinHash and SimHash can miss; Vicinity abstracts across competing approximate-nearest-neighbor backends; and Semble competes with grep-plus-file-reading and code-specialized transformer retrieval while emphasizing CPU speed, low token use, and no external services.
Leadership
Founders
Thomas van Dongen
Co-founder of Minish and head of AI engineering at Springer Nature, where he leads production AI systems for scientific publishing, including research discovery, compliance screening, and recommender systems. His public GitHub profile describes him as an open-source ML engineer working on efficient NLP models and packages.
Stéphan Tulkens
Co-founder of Minish and a computational linguistics/AI specialist. He earned a PhD at the University of Antwerp's CLiPS group on orthographic processing in reading, previously worked at Slimmer AI, and is a machine-learning engineer at Ecosia.
Executive Team
Thomas van Dongen
Co-founder
Head of AI engineering at Springer Nature and production ML engineer working on embeddings, retrieval, language models, and AI tools for scientific publishing.
Stéphan Tulkens
Co-founder
Computational linguistics and AI researcher with a University of Antwerp PhD; machine-learning engineer at Ecosia and former Slimmer AI employee.
Founding Story
Minish began as the collaboration of Thomas van Dongen and Stéphan Tulkens to make NLP models fast and small. The initial Model2Vec idea started as a possible cache for sentence transformers; the founders found they could instead store fixed token representations and average them at inference time, enabling highly efficient static embeddings without requiring a corpus for basic distillation.
Business Model
Revenue Model
Minish presents itself as a free/open-source lab. Its public products are distributed as open-source Python packages, Rust crates, GitHub repositories and Hugging Face models; no paid SaaS, API or license pricing is described on the official pages.
Target Markets
- NLP and machine-learning engineers
- Developers building search, RAG, classification, and retrieval systems
- Teams curating or training large language-model datasets
- Companies needing CPU, edge, mobile, or low-cost inference
- AI coding-agent and developer-tool builders
- Researchers evaluating embedding and nearest-neighbor methods
- Text classification, clustering, semantic search, and retrieval-augmented generation
- Embedding large corpora such as English Wikipedia on commodity hardware
- Dataset deduplication, train/test overlap checks, outlier filtering, and corpus inspection
- Lightweight vector search and approximate-nearest-neighbor evaluation
- Edge and mobile NLP deployment where memory and storage are limited
- Natural-language code search for coding agents and MCP-compatible development tools