Schematron V2 Turbo

1. Throughput is the main reason Turbo exists

This is where Turbo separates itself.

On a single H100, with a test setup of 10K input tokens and 500 output tokens, Turbo reaches:

4.14 requests/sec

For comparison:

• V2 Small: 2.47 requests/sec
• First-generation 8B model: around 1.66 requests/sec

Turbo is about 1.7× faster than V2 Small and around 2.5× faster than the original 8B model.

For a normal user, this difference may not matter.

For systems that constantly process pages, prices, or financial documents, throughput affects everything from waiting time to server cost.

If I were choosing where to use Turbo, I would not start with a small number of important pages.

I would put it into high-volume, repetitive workloads where speed actually changes the economics.

2. Good enough quality, not the highest quality

Turbo is not trying to beat Small on accuracy.

It is making a different trade-off.

Inference.net evaluated the models with GPT-5.4 as an LLM judge (1–5 score):

• V2 Turbo: 4.039
• V2 Small: 4.060
• V1 8B: 4.070
• V1 3B: 3.909

SimpleQA results:

• V2 Turbo + GPT-5 Nano + Exa: 79.42
• V2 Small: 83.10

The numbers show Turbo is slightly behind Small.

Personally, I would not focus too much on the decimal difference here.

For a production extraction pipeline, the bigger questions are:

• How often does it fail?
• How much cleanup is needed?
• How much does each page cost?

If the output already meets the business requirement, paying more for a slightly higher score may not make sense.

3. Schema-driven, not prompt-driven

Like V2 Small, Turbo is controlled through Schema.

You do not tell it:

“Extract the title.”

You define the structure first:

{
  "title": {
    "type": "string"
  }
}

Then the model fills that structure from the HTML.

The benefit is predictable output.

You spend less time dealing with:

• missing fields
• wrong data types
• broken JSON formatting

The limitation is also clear.

When the page structure changes, the schema needs to change.

Turbo is not a webpage assistant that understands everything on a page. It is a structured extraction tool.

4. Compared with V1, Turbo is the faster successor to the original 3B model

Inference.net has retired the original 3B and 8B versions.

Requests using schematron-3b are automatically routed to V2 Turbo.

The upgrade is straightforward:

Speed:

• V1 3B: around 1.66 requests/sec
• V2 Turbo: 4.14 requests/sec

Quality:

• V1 3B: 3.909
• V2 Turbo: 4.039

At the same time, pricing is lower.

For teams already using the original 3B model, the upgrade path is obvious.

Turbo is not a smaller version of Small.

It is built for a different workload: moving large amounts of structured data quickly.

I would use Schematron V2 Turbo for:

• large-scale web scraping
• product price monitoring
• bulk financial document extraction
• HTML-to-database pipelines

These jobs need stable extraction, not a model that explains what a webpage means.

I would not use it for:

• content summaries
• page interpretation
• complex analysis
• open-ended questions

If a system processes hundreds of thousands or millions of pages, Turbo’s speed and pricing advantage are meaningful.

If I only need to extract information from a few pages occasionally, a general-purpose model is probably easier.

My decision rule would be simple:

Does Turbo reduce waiting time and manual cleanup in the pipeline?

If yes, it earns a place.

If the cheap API price comes with too much manual checking, the savings disappear.