What it is
Chonkie does one job in a retrieval-augmented generation (RAG) pipeline: it cuts already-extracted text into chunks before they are embedded and stored. Version 1.7.0 lists eleven chunkers, from fixed-size token and sentence splitting through recursive, semantic, late, code, table and neural chunking to an LLM-driven "Slumber" chunker, plus a Pipeline API and a self-hosted REST server.
In August 2026 the maintainers announced that 1.7 will be the last 1.x release. Chonkie v2 is a rewrite as a small Rust core with Python and JavaScript bindings that keeps the token, sentence, recursive and code chunkers and removes the vector-database integrations and other plumbing; the main branch is frozen to pull requests meanwhile. The hosted Chonkie Cloud no longer resolves.
It is MIT-licensed and built by Chonkie, Inc., a Y Combinator company that now trades as Feyn and sells custom model training; chonkie.ai redirects to usefeyn.com and the repository lives in the feyninc organisation.
Our review
Chonkie is a well-made answer to a question that matters less than people think, and its own maintainers now agree. In the August 2026 v2 announcement they report that, across hundreds of runs on their MTCB benchmark, the semantic, neural and Slumber chunkers were either much worse than the recursive chunker or within 2 to 3 per cent of it. That matches outside work: a 2024 Vectara study found semantic chunking's extra compute was not justified, and an August 2026 study of eight strategies found the winner shifts with the embedding model, dataset and metric.
The library is popular. The announcement cites 1.3 million downloads a month, and the repo has about 4,800 stars. The base install is small (the README claims a 505KB wheel) and the recursive chunker with its Markdown recipe is a sound default. The speed claims, such as a SIMD chunker at "100+ GB/s", are the project's own benchmarks.
The problem is timing. v1.7.0 shipped on 7 July 2026 and is the last 1.x release; fixes go to v2, the main branch is frozen to pull requests, and the last commit on main was 26 August 2026. v2 drops the integrations (the "handshakes" to Qdrant, pgvector and others), the chefs and the stale cloud chunkers, and promises a 1.x-to-2.0 migration guide. Open bugs include a CodeChunker crash with tree-sitter-language-pack 1.16.1 and string tokenisers that ignore HF_HUB_OFFLINE and block on the network.
A team adopting it now is buying a dependency it will have to migrate within months, onto an API it cannot yet read. That is a reasonable bet for someone who only needs recursive or code chunking, which v2 keeps. It is a poor one for anyone choosing Chonkie for its menu of strategies or its vector-store connectors, which LlamaIndex and LangChain splitters already cover with longer support horizons.
Where it earns its place
A platform engineer running nightly ingestion of 200,000 Markdown help-centre pages
Split documents consistently before embedding without pulling in a whole RAG framework
pip install chonkie and use RecursiveChunker with the markdown recipe; the small install suits a serverless batch job. Pin 1.7.0 and plan a v2 migration, since no further 1.x fixes are coming.
An ML engineer at a legal-tech start-up evaluating retrieval quality
Test whether semantic or late chunking beats recursive chunking on their contracts
Swap chunkers behind one API and score recall on a held-out question set. Worth a day, not a quarter: the maintainers found fancy chunkers within 2 to 3 per cent of recursive, and v2 drops them.
A head of data choosing an ingestion stack for a two-year product
Pick dependencies that will still be supported next year
Treat Chonkie as a chunking primitive only. Its vector-store handshakes are being removed in v2, so wire Qdrant or pgvector directly, or use LlamaIndex if you want connectors maintained for you.
Before you commit
- Setup
- pip install chonkie on Python 3.10 or newer; heavier chunkers need extras such as chonkie[semantic] or chonkie[code]. A first chunked document takes minutes on a laptop CPU.
- Integrations
- v1.7 ships handshakes for Chroma, Elasticsearch, MongoDB, pgvector, Pinecone, Qdrant and others, a Mistral OCR chef and a REST server; v2 will remove the handshakes and chefs.
- Data and privacy
- Runs locally with no telemetry found in the code. String tokeniser names are fetched from the Hugging Face Hub, and an open issue reports this ignores HF_HUB_OFFLINE. The REST server stores pipelines in local SQLite.
- Cost at scale
- Free under MIT; recursive and token chunking run on CPU. Costs come from the embedding or LLM calls behind semantic, late or Slumber chunking. The former Chonkie Cloud no longer resolves and Feyn sells no chunking product.
- Limits
- 49 open issues and pull requests, main frozen during the v2 rewrite; it does not parse PDFs or Office files, so Docling or similar must sit in front.
What's good
- Small core install with optional extras
- Recursive and code chunkers that survive into v2
- Maintainers publish candid benchmark findings
- Permissive MIT licence
What's not
- v1.7 is the last 1.x release; main is frozen
- v2 removes semantic, neural and Slumber chunkers and all integrations
- Company focus has moved to custom model training
Best for
Teams that need a light recursive or code chunking step and can accept a v2 migration soon.
Alternatives
Chonkie: common questions
- Is Chonkie worth it?
- A tidy chunking library mid-rewrite: v1.7 works today, but most of what makes it distinctive is scheduled for removal in v2. We rate it 3.7 out of 5.
- How much does Chonkie cost?
- Free and open source under the MIT licence.
- What are the downsides of Chonkie?
- v1.7 is the last 1.x release; main is frozen; v2 removes semantic, neural and Slumber chunkers and all integrations; company focus has moved to custom model training.
- Who is Chonkie best for?
- Teams that need a light recursive or code chunking step and can accept a v2 migration soon.
- How much does Chonkie cost for a team or in production?
- Free under MIT; recursive and token chunking run on CPU. Costs come from the embedding or LLM calls behind semantic, late or Slumber chunking. The former Chonkie Cloud no longer resolves and Feyn sells no chunking product.
- Is Chonkie safe for company data?
- Runs locally with no telemetry found in the code. String tokeniser names are fetched from the Hugging Face Hub, and an open issue reports this ignores HF_HUB_OFFLINE. The REST server stores pipelines in local SQLite.
- What are the best Chonkie alternatives?
- The closest alternatives we have reviewed are LlamaIndex, LangChain and Docling.
Answers checked Oct 23, 2026.