Inference.net Releases Schematron V2 Small, a 3B-Parameter Model for HTML-to-JSON Extraction

Inference.net has released Schematron V2 Small, a 3B-parameter model purpose-built for HTML-to-JSON data extraction. The model is available now through OpenRouter and Inference.net's own API.

What It Does

Schematron V2 Small takes raw HTML input and converts it into structured JSON output according to a user-defined schema. According to Inference.net, the model prioritizes extraction accuracy on complex schemas and long pages — a common bottleneck for scraping pipelines, data engineering workflows, and agents that need to parse web content into machine-readable formats.

The model has one notable usage constraint: extraction instructions must be supplied through a JSON schema passed in the response_format parameter, not through system or user prompts. This differs from how most general-purpose LLMs handle structured output requests, and integrators will need to adjust their calling conventions accordingly.

Specs and Pricing

• Parameters: 3B
• Context window: 128,000 tokens
• Input price: $0.05 per 1M tokens
• Output price: $0.23 per 1M tokens
• P50 latency: 1.01 seconds (best provider)
• Throughput: 58 tokens per second
• Release date: September 12, 2026

At 3B parameters, Schematron V2 Small is designed for speed and cost efficiency rather than general reasoning. Inference.net has not published independent benchmark scores for the model on OpenRouter's listing, and no MMLU, HumanEval, or comparable extraction-accuracy figures were disclosed.

Availability

The model is served through Inference.net as the sole listed provider on OpenRouter at launch. OpenRouter notes that uptime and app-usage data are not yet available, reflecting the model's recent release. Latency and throughput figures reflect data collected over the past 24 hours across available regions.

What This Means

Schematron V2 Small is a narrow, task-specific model rather than a general-purpose competitor to frontier LLMs. Its value proposition is cost and speed for a single job — turning messy HTML into clean JSON — at a price point ($0.05/$0.23 per 1M tokens) far below general-purpose models capable of the same task. For teams running high-volume scraping or data-extraction pipelines, a specialized small model can cut costs significantly compared to routing every extraction call through a large frontier model.

The requirement to pass schemas via response_format rather than prompts signals that Inference.net is optimizing for programmatic, API-driven use rather than conversational or ad hoc extraction. This is consistent with a broader trend of vendors releasing smaller, task-specific models rather than competing solely on general capability — a strategy that trades flexibility for lower latency and cost at scale. Whether Schematron V2 Small holds up against general-purpose models on real-world extraction accuracy remains unverified, as no independent benchmark data accompanies this release.