Inference.net Launches Schematron V2 Turbo, a 3B-Parameter Model for High-Volume HTML-to-JSON Extraction

Inference.net has released Schematron V2 Turbo, a 3-billion-parameter model purpose-built for extracting structured JSON data from raw HTML at scale. The model is now available through OpenRouter and Inference.net's own API.

Unlike general-purpose language models, Schematron V2 Turbo is narrowly scoped: it converts HTML input into JSON output according to a schema supplied via the response_format parameter. Extraction instructions cannot be passed through system or user prompts — the schema itself defines what gets extracted. This constrained design is aimed squarely at pipeline and infrastructure use cases rather than conversational tasks.

Pricing and specs

Schematron V2 Turbo is priced at $0.03 per 1M input tokens and $0.15 per 1M output tokens, and supports a 128K token context window. According to OpenRouter's listing, the model's best-observed performance from the inference.net endpoint runs at a P50 latency of 4.48 seconds and throughput of 26 tokens per second.

The model carries a listed release date of September 12, 2026, though it is already live and accessible through OpenRouter's model catalog and playground.

Built for throughput, not versatility

Inference.net describes Schematron V2 Turbo as optimized for "high-volume extraction workloads," which explains both its small parameter count and its restrictive input format. At 3B parameters, the model is far smaller than general-purpose frontier models, trading broad capability for speed and cost efficiency on a single, narrow task: turning messy or semi-structured HTML into clean JSON that conforms to a caller-defined schema.

This positions Schematron V2 Turbo as a component for data pipelines — web scraping, content extraction, document processing — rather than as a chatbot or coding assistant. Its schema-driven interface also makes it easy to slot into existing structured-output workflows without prompt engineering.

No benchmark scores, training data cutoff, or architectural details beyond parameter count have been disclosed by Inference.net at this time.

What this means

Schematron V2 Turbo is a specialist tool, not a general model, and its value proposition rests entirely on throughput and cost at scale rather than raw capability. A 3B-parameter model priced at fractions of a cent per request makes sense for teams running millions of extraction calls against web content, where a frontier model's cost and latency would be prohibitive. The tradeoff is inflexibility: because instructions must come through a JSON schema rather than natural language prompts, the model is less adaptable to edge cases than a general-purpose LLM instructed via prompt. For teams with well-defined, repetitive extraction tasks, that's likely an acceptable trade. For anything requiring nuanced judgment about what to extract, it isn't.