1. Schema-driven extraction
The biggest difference between Schematron and general AI models is that it follows a schema instead of a prompt.
You do not ask:
“Find the product information.”
You define exactly what you want:
- product name
- price
- category
- attributes
The model then maps webpage content into those fields.
For large-scale data pipelines, this approach is much easier to manage than calling a general model and fixing the output afterward.
The limitation is that there is not much room for interpretation.
If the structure of the website changes, your schema needs to change too.
Schematron is better viewed as a structured extraction tool, not a webpage analyst.
2. Long-page handling
Schematron V2 Small supports a 128K token context window and is designed to handle long HTML pages.
Inference.net says the model is trained to deal with noisy webpage markup and messy real-world HTML.
That part matters more to me than a small benchmark improvement.
Real webpages are rarely clean. They contain ads, scripts, nested elements, and unnecessary content.
If a model can handle that without heavy preprocessing, it saves practical engineering time.
The company still recommends removing elements like scripts and styles with tools such as lxml.
3. Small model, low cost, and fast processing
Schematron V2 Small has only 3B parameters.
That is a major reason it can stay inexpensive.
Inference.net reports a throughput of 2.47 requests per second on a single H100, using a test setup with 10K input tokens and 500 output tokens.
It is not designed to compete on complex reasoning speed.
That is not the point.
Its job is repetitive processing: receive HTML, extract fields, return structured data.
Benchable testing places it in the 93rd percentile for speed, with 100% reliability in its reported evaluation.
I care less about the ranking itself and more about production behavior.
For an extraction model, a small number of formatting failures across thousands of pages can create more work than a slightly slower response.
4. V2 Small closes the gap with the larger V1 model
Schematron V1 had both 3B and 8B versions.
The goal of V2 Small is to get close to the 8B version while keeping the smaller model’s efficiency.
Inference.net used GPT-5.4 as an LLM judge:
- V2 Small: 4.060
- V1 8B: 4.070
- V1 3B: 3.909
On SimpleQA:
- V2 Small + GPT-5 Nano + Exa: 83.10
- V1 8B: 85.58
The 8B model still scores slightly higher.
But I am not sure that difference matters much in production.
If a system is processing hundreds of thousands of webpages, cost, speed, and consistency usually matter more than a small benchmark gap.
That is where a smaller model can make more sense.
Comments (0)