Open-source Chinese synonym tool that put its model behind a paywall

What it does

Synonyms is a Python toolkit that wraps a large, pre-trained Chinese word2vec model to serve up near-synonyms, sentence similarity, keyword extraction, and word segmentation. It targets classic NLP chores—RAG retrieval, recommendation logic, search ranking, and summarization—where you need to know that “飞机” and “直升机” are closer than “飞机” and “命运”. The library downloads a 435,729-term vector file on first use and keeps lookups local after that.

The interesting bit

The project recently pivoted from a permissive open-source license to the “Chunsong License,” and now requires a paid license ID to download the underlying model file. That makes it one of the rarer cases where the code remains freely visible on GitHub but the weights are treated as commercial software. The README explicitly notes that prior contributors can contact the company to “discuss the charging issue,” which is an unusual level of frankness about monetizing a community project.

Key highlights

• Vocabulary of 435,729 Chinese terms—substantially larger than the Tongyici Cilin or HowNet baselines cited in the docs.
• Fast local lookups: benchmark claims ~0.2 microseconds per nearby() call on modest hardware.
• Bundles several utility APIs: sentence similarity (compare), keyword extraction (keywords), segmentation (seg), and raw vector access (v, sv, bow).
• Supports swapping in custom word2vec binaries and segmentation dictionaries via environment variables.
• Data derived from a Wikidata corpus rather than general web crawl, though the README does not detail the training pipeline.

Caveats

• The core model is gated behind a commercial license from the Chatopera store; the weights are no longer freely redistributable as of October 2023.
• Segmentation does not remove stopwords or punctuation, so downstream tasks must handle cleanup themselves.
• The README warns that similarity scores and benchmarks may shift as the model is retrained, so pinned reproducibility is not guaranteed.

Verdict

Grab it if you need battle-tested Chinese similarity APIs and your budget covers a model license. Skip it if you are looking for a fully gratis, open-weights alternative or need transparent training data provenance.