Free and open source · Product · ReDefiNER
Pull out the details you need. On your own servers.
ReDefiNER is a 0.8-billion-parameter model that finds and labels the details you define in English text, such as names, account numbers, products or order IDs. You describe them in a plain-English rulebook. It runs on an ordinary CPU, works offline, and is free for commercial use under Apache 2.0.
Same message. Two rulebooks.
"Hi, this is Priya Raman (account 88-41207). Since updating Zoom to 6.2.5 on my MacBook Air, calls drop with error 1132. Reach me at priya.raman@outlook.com or 617-555-0148."
Privacy team asks for person, account, email, phone
Product team asks for product, version, device, error
The same model, nothing retrained. Only the rulebook changed.
How it works
Three steps. No retraining.
Two modes: extract finds and labels every matching detail; label types the spans you supply.
Write a rulebook
List the kinds of details you want, give each a one-line description, and add rules in plain English if you need them.
[TYPE] PERSON [DESCRIPTION] A person's name. PE1 Leave out titles such as "Dr".
Give it your text
Emails, support tickets, forms, CVs, reports or search queries. Send one text or thousands. Long documents are split up and read in pieces for you.
Get exact answers back
Each answer is a piece of your own text, with where it sits, its type and a confidence score. Nothing is generated, so it can't add words that aren't in your document.
Real output
Add one rule. See the answer change.
From the released model on a laptop CPU. The only thing that changed is the rule.
Keep "under" with the price
"red nike air max size 10 under $120"
PR1 A price limit keeps its limit word in the span: "under $50".
Skip your own company
"Hi Acme team, my Sonos Arc stopped pairing after the Acme Home update. Sonos says it is an Acme problem."
NE1 Our own company and products: "Acme", "Acme Home". Never label them.
Let the format decide
"Where is my order INV-7731? The invoice for it, #D-55102, says it shipped."
A reference starting with # is ORDER_ID. One starting with INV- is INVOICE_ID.
Why teams use it
Strong accuracy for its size, at a fraction of the cost.
A good fit for hospitals, banks, schools, government and anyone who can't send their data out.
Strong accuracy for its size
On a standard extraction test it scores 68.1, about 5 points behind Claude Haiku 5.5. On 13 public test sets it also beats a research model about 40 times its size.
A fraction of the cost
No GPU and no per-use fees. In our test, 1,000 requests cost about 2 cents on a small CPU server, against 19 cents on Claude Haiku 5.5 and $2.42 on Claude Opus 5.5.
Private and on-premise
It runs on your own servers and works offline, so nothing goes to an outside AI service.
Fast
About 0.05 seconds per request on a GPU, against 2 to 17 seconds for the frontier AI services in our test.
Free for commercial use
The code and the model are under Apache 2.0. Every benchmark result is public too, including where it loses.
You stay in control
Change what it finds by editing plain-English rules. No retraining, no labelled data, no machine learning team.
Benchmarks
Small model. Big savings.
Five frontier models got the same rulebook and the same test. They score higher, by 4.3 to 11.3 points on extraction. ReDefiNER costs far less, answers faster and runs on your own servers.
| Model | Accuracy (F1) | Cost per 1,000 requests | Seconds per request |
|---|---|---|---|
| Claude Opus 5.5 | 79.4 | $2.42 | 2.2 |
| Claude Opus 4.8 | 77.1 | $2.92 | 4.2 |
| GPT-5.4 | 73.5 | $5.32 | 8.6 |
| Claude Haiku 5.5 | 73.1 | $0.19 | 2.3 |
| Gemini 3.1 Pro | 72.4 | $17.27 | 16.8 |
| ReDefiNER, on a GPU | 68.1 | $0.002 | 0.05 |
| ReDefiNER, on a 2-CPU server | 68.1 | $0.02 | 1.3 |
CrossNER AI extraction test, the same rulebook for every model, run on 1 November 2026, with the rulebook cached. F1 is the standard accuracy score for this task; higher is better. See the full comparison and method.
Where it fits
The same details, out of a lot of text.
Ready-made rulebooks are included for several of these.
Customer support
Order numbers, products, versions and error codes from every ticket.
Personal details
Names, phone numbers, emails and addresses, to help find what to hide before sharing a document.
CVs and job posts
Skills, tools and programming languages, sorted the way you define them.
Shop search boxes
Turn "red nike air max size 10 under $120" into brand, product, colour, size and price.
Research papers
Methods, datasets, metrics and tasks from abstracts.
Travel and bookings
Who is booking and who is flying, where from, where to, and which date is the return.
Choosing
A big AI service, or ReDefiNER?
For pulling the same kinds of details out of lots of text, the trade often makes sense. Keep the big models for open-ended reasoning and writing.
| Frontier AI model, online | ReDefiNER | |
|---|---|---|
| Accuracy at pulling details out of text | Higher: 72 to 79 in our test | A few points lower: 68 in our test |
| Where your text goes | To the provider's servers | Stays on your own machine |
| Cost | You pay for every request: $0.19 to $17 per 1,000 in our test | Free model. You pay only for your server: about $0.002 to $0.02 per 1,000 |
| Time per request | 2 to 17 seconds | 0.05 seconds on a GPU |
| Hardware | The provider's GPUs | An ordinary CPU is enough |
| What you get back | Generated text, which can include words that aren't in your document | Only exact pieces of your text, with their positions |
| Best for | Open-ended reasoning and writing | High-volume, repeatable extraction |
Get started
Install it and run it in minutes.
Python 3.10 or later. Works on a CPU or an NVIDIA GPU, and offline once downloaded.
pip install redefinerOr run it as a local server
redefiner serve --model jungroo-ai/redefiner-0.8b --port 8000
from redefiner import Tagger
tagger = Tagger.from_pretrained("jungroo-ai/redefiner-0.8b")
rulebook = """[RULES]
Extract every entity of the supplied types. Return exact spans from the text.
[TYPE] PERSON
[DESCRIPTION] A person's name.
PE1 Label the full name; leave out titles such as "Dr" or "Prof".
[TYPE] ORGANIZATION
[DESCRIPTION] A company, university or other institution.
"""
tagger.extract("Prof Ada Lovelace corresponded with the Royal Society.", rulebook)
# [{'start': 5, 'end': 17, 'text': 'Ada Lovelace', 'label': 'PERSON', 'confidence': 0.977},
# {'start': 40, 'end': 53, 'text': 'Royal Society', 'label': 'ORGANIZATION', 'confidence': 0.489}]At a glance
The details.
Release 1.0.7. The model is built on Qwen3.5 0.8B, which is also Apache 2.0.
| Does | Extract mode finds and labels spans; label mode types spans you supply. English text. |
| Controlled by | A plain-English rulebook of types, one-line descriptions and rules. No retraining. |
| Returns | Exact character spans, each with a type and a confidence score. Nothing is generated. |
| Size | 0.8 billion parameters, one 0.77 GB int8 file. |
| Hardware | A CPU with about 2 to 2.6 GB of memory at peak, or an NVIDIA GPU. Works offline once downloaded. |
| Interfaces | Python (extract and label) and a local HTTP server. |
| Limits | English only; spans up to 16 tokens; rulebooks tested up to about 10,000 tokens; long texts are read in parts, so labels can differ across a document. |
| Licence | Apache 2.0: free to use, change and share, including in commercial products. Keep the licence and copyright notices. |
The model can miss or mislabel things, so check its output before acting on it in legal, medical or financial work, and don't use it as the only safeguard when finding personal data to redact.
Built by Jungroo AI Labs
Want help putting it to work?
We designed, trained and tested ReDefiNER ourselves, and published every result. We can write and test rulebooks on your real data, run it privately on your cloud or your own servers, and connect it to your email, ticketing or document stores. For a model trained on your own documents, see Jungroo AI Studio.
Questions
Frequently asked.
What is ReDefiNER?
ReDefiNER is a free, open-source 0.8-billion-parameter model from Jungroo AI Labs that finds and labels the kinds of details you define in English text, such as names, account numbers, products or order IDs. You describe them in a plain-English rulebook, with no retraining.
Is ReDefiNER free for commercial use?
Yes. The code and the model are under the Apache 2.0 licence, so you can use, change and build them into commercial products, keeping the licence and copyright notices.
Does ReDefiNER need a GPU?
No. It runs on an ordinary CPU with about 2 to 2.6 GB of memory and works offline. On an NVIDIA GPU it answers in about 0.05 seconds per request.
How accurate is ReDefiNER compared with large AI models?
In a published test on 1 November 2026, five frontier models scored 4.3 to 11.3 points higher on extraction (F1). ReDefiNER scored 68.1, about 5 points behind Claude Haiku 5.5, while costing far less per request, answering faster and running on your own servers.
Which languages does ReDefiNER support?
English only, for now.
Contact
Deploy with precision.
We partner with a limited number of enterprises at a time, so every engagement gets senior attention. Brief our leadership on your roadmap.