Inference.net: Schematron V2 Turbo

Schematron V2 Turbo is a text model from Inference.net, listed on api14 under the model id inference-net/schematron-v2-turbo. It is a 3B-parameter HTML-to-JSON extraction model, and Inference.net positions it as a throughput-first option for high-volume extraction workloads. Its provider release date is 2026-09-12. If you are building a pipeline that turns web pages or HTML fragments into structured records, this is the model family it belongs to.

The model takes text as input and returns text as output, with a context window of 128,000 tokens and a maximum output of 8,192 tokens. Its defining behavior is that extraction instructions must be supplied through a JSON schema in response_format rather than in free-form prose. That constraint shapes how you prompt it: instead of describing the fields you want in natural language, you declare them as a schema, and the model fills that schema from the HTML it is given. The large context window means you can pass substantial documents in a single request, while the 8,192-token output ceiling bounds how much structured data one call can return.

The practical use cases follow directly from that design. High-volume HTML-to-JSON extraction is the primary one: scraping product listings, parsing invoices or receipts rendered as markup, pulling fields out of listing pages, or normalizing semi-structured documents into records your application can store. Because the model is small and tuned for throughput, it suits batch jobs where you run the same schema across thousands of pages and care more about consistent, schema-conformant output than about open-ended reasoning. It is a text-only model, so it does not accept images or audio, and it does not generate anything other than text.

You call it through api14's OpenAI-compatible API using the model id inference-net/schematron-v2-turbo. Because the interface is OpenAI-compatible, existing OpenAI SDKs and clients work with a change of base URL and model id. The supported request parameters are frequency_penalty, logit_bias, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, top_k and top_p. In practice you will lean on response_format and structured_outputs to pass your JSON schema, max_tokens to cap the returned record, and seed or temperature when you want reproducible extraction across a batch.

The main thing to plan around is the schema requirement. If your task is open-ended generation, summarization, or conversational reasoning rather than extraction into a fixed shape, a general-purpose text model will fit better. Likewise, if you need image or audio input, this model does not support it. Within its intended niche, though, the combination of a 128,000-token context window, an 8,192-token output limit, and schema-driven output makes it a reasonable choice for large-scale HTML-to-JSON jobs where throughput matters.