Datalab Releases LIFT, a 9B Open-Weights Vision Model That Reads PDFs into JSON
Datalab has released LIFT, a 9B open-weights vision model that turns messy PDFs into clean, structured data. In plain words: you give LIFT a PDF and a simple template, and it gives you back neat, organised information a computer can use. A “vision model” is an AI that can look at images and documents, not just text. “Open-weights” means the model’s trained brain is shared publicly, so developers can download and run it themselves. The “9B” means it has 9 billion parameters, which are the internal settings that make the model work. More parameters usually means a smarter model.
This matters because reading PDFs has always been a pain. Invoices, forms, and reports come in endless formats. LIFT promises to pull the exact fields you want, every time, in a format that just works.
What does LIFT actually do?
LIFT extracts structured JSON from PDFs using schemas. Let us unpack that. JSON is a simple, tidy way to store data that software can read easily, like a labelled list of values. A “schema” is a template that tells the model exactly what fields you want, for example invoice number, date, and total amount.
So you hand LIFT a document and a schema. It reads the document in a single pass and fills in your template. It can even handle multi-page PDFs where one value is spread across several pages. If a field is missing in the document, LIFT can return “null”, meaning empty, instead of making something up. This trained ability to say “not found” is called abstention, and it is very useful for accuracy.
How it works under the hood
LIFT uses a clever trick called schema-constrained decoding. It turns your JSON schema into a strict set of rules. As the model writes its answer word by word, the system blocks any output that would break the template. The result is that LIFT’s answer always matches the structure you asked for. You never get broken or messy output. This is a big deal for developers who need reliable data, not surprises.
Benchmarks and specs
Datalab shared test scores for LIFT. A “benchmark” is a standard test used to compare AI models fairly. These figures are as reported by Datalab on a 225-document test set. Two scores matter here. “Field accuracy” is how often LIFT gets a single field right. “Full-document accuracy” is how often it gets every single field in a document right, which is a much harder bar.
| Model | Size | Field accuracy | Full-doc accuracy | Median latency |
|---|---|---|---|---|
| LIFT (Datalab) | 9B | 90.2% | 20.9% | 9.5s |
| Qwen3.5-9B | 9B | 76.32% | 24.0% | 16.8s |
| NuExtract3 | 4B | 81.5% | 8.4% | 8.3s |
| Gemini Flash 3.5 | — | 91.3% | 40.0% | 28.1s |
| Datalab API (cloud) | — | 95.9% | 44.4% | 30.8s |
What it means: among models you can download and run yourself, LIFT leads on field accuracy at 90.2%. It beats the similar-sized Qwen3.5-9B on accuracy and speed. “Latency” is the time taken per document; lower is faster. LIFT is quick at 9.5 seconds median per document when handling 8 requests at once. The big cloud options, Gemini Flash 3.5 and Datalab’s own API, still score higher on full-document accuracy, but they are slower and are not open-weights. So LIFT is the strong free, self-hosted choice.
License, price, and access
Here is the practical part. The code uses the Apache 2.0 license, a very open and friendly license. The weights use a modified OpenRAIL-M license. A “license” is the legal rulebook for how you may use the software. LIFT is free for research, personal use, and startups with under $5 million in funding or revenue. Bigger companies will likely need a commercial arrangement.
You can get LIFT from GitHub and HuggingFace, try it in a Playground, or use the hosted API. Installation is a single command: pip install lift-pdf. It runs on common GPUs, including the H100, A100, L40S, A10, L4, and even older cards like the 4090, 3090, and T4. A GPU is a chip that runs AI fast.
| Spec | Detail |
|---|---|
| Model | LIFT by Datalab |
| Size | 9 billion parameters |
| Type | Open-weights vision-language model |
| Input | PDFs and images + a JSON schema |
| Output | Structured JSON |
| Code license | Apache 2.0 |
| Weights license | Modified OpenRAIL-M |
| Free for | Research, personal use, startups under $5M |
| Install | pip install lift-pdf |
| Released | June 2026 |
Who is Datalab?
Datalab is not new to this space. The same team built popular open-source document tools called chandra, marker, and surya. These are well-known for OCR and document processing. OCR, or Optical Character Recognition, is the technology that reads text from images. With LIFT, the team moves a step further, from just reading text to understanding and organising it.
FAQ
What is LIFT in one sentence?
LIFT is a free, downloadable 9-billion-parameter AI model that reads PDFs and images and returns the exact data fields you ask for as clean, structured JSON.
Is LIFT free to use?
Yes, for research, personal use, and startups with under $5 million in funding or revenue. The code is Apache 2.0 and the weights use a modified OpenRAIL-M license. Larger firms may need a commercial deal.
How is LIFT different from normal OCR?
OCR just reads text off an image. LIFT goes further. It understands the document and pulls out the specific fields you want, in a structured format, even handling missing values and multi-page documents.
Why it matters (especially for India / founders)
India runs on documents. Invoices, KYC forms, bank statements, GST filings, and government records all need to be read and entered into systems. Today, much of this is done by hand. A free, accurate model like LIFT could automate huge amounts of this work.
For founders, the open-weights angle is the gift. Because LIFT is free for small startups and runs on cheap GPUs, an Indian fintech or SaaS team can build document automation without paying per-page fees to a big cloud provider. This is part of a wider shift toward smart, specialised AI tools, much like the rise of new orchestration models that route work between AIs. The flip side is security: if you process sensitive documents, guard that data carefully, a lesson underlined by incidents like the recent LastPass data theft.
The takeaway
LIFT is a smart, practical release. It is the leading open-weights model for turning PDFs into structured data, with strong field accuracy, fast speeds, and a license that welcomes small startups. It will not beat the biggest cloud models on every score, but it does not need to. For developers who want to own their pipeline and skip per-page fees, LIFT is a powerful new tool, and a clear win for anyone drowning in PDFs.
