Lab release history
Last updated Sep 12, 2026
Inference.net model releases
US AI infrastructure startup running a distributed GPU inference network, positioned on low-cost, high-throughput serving of small purpose-built models rather than frontier scale. Ships the "workhorse model" line — compact task-specific LLMs sold on cost per unit of work — of which the Schematron family (HTML-to-JSON structured extraction) is the flagship. Models are served API-only through its own platform plus OpenRouter. This page collects the lab's model releases, lifecycle events, source links, and model metadata in one crawlable record.
Most recent in this set
2 models
Schematron V2 Small
AvailableThe quality-oriented half of Inference.net's Schematron V2 pair, listed Sep 12 2026 (inference-net/schematron-v2-small). Also a 3B-parameter HTML-to-JSON extraction model, but tuned to hold up on complex schemas and long pages where the throughput-optimized Turbo variant degrades. Like Turbo it is schema-driven — the extraction target goes in a JSON schema via response_format rather than the prompt — with a 128K-token context and 4,096 max output tokens, text in and text out. Proprietary and API-only via Inference.net and OpenRouter, at $0.05 input / $0.23 output per Mtok with cache reads at $0.05; OpenRouter reported ~1.74s P50 latency at listing. Inference.net reports an LLM-as-judge quality score of 4.060 and 83.10 on SimpleQA, both slightly ahead of Turbo's 4.039 / 79.42 — the trade the two variants are meant to express. In scope as a narrow document/structured-extraction LLM, consistent with the Cohere Parse 5 and North Micro Vision precedent.
Schematron V2 Turbo
AvailableA 3B-parameter HTML-to-JSON extraction model from Inference.net, listed Sep 12 2026 (inference-net/schematron-v2-turbo). Part of the company's "workhorse model" line — small, purpose-built LLMs sold on cost per unit of work rather than general capability — and tuned for maximum throughput on high-volume scraping and ingestion pipelines, reported at ~4.14 requests/second on a single H100. It turns messy HTML into clean structured JSON, and is schema-driven rather than prompt-driven: the extraction target is supplied as a JSON schema in the response_format parameter, not in the system or user prompt. 128K-token context, 8,192 max output tokens, text in and text out. Proprietary and API-only via Inference.net and OpenRouter, at $0.03 input / $0.15 output per Mtok. Inference.net reports an LLM-as-judge quality score of 4.039 and 79.42 on SimpleQA — positioning it as the cheaper, faster half of the V2 pair against the higher-quality Small. In scope as a narrow document/structured-extraction LLM, consistent with the Cohere Parse 5 and North Micro Vision precedent.