April 2026 Summaries

Schematron V2 is the latest iteration of specialized HTML-to-JSON extraction models, offering enhanced performance with two new variants, Schematron V2 Small and Schematron V2 Turbo, which significantly improve upon the previous generation's speed and quality. Optimized for cost and latency, these models maintain high extraction quality while providing faster processing capabilities, making them suitable for large-scale data extraction tasks. Schematron V2 Small nearly matches the quality of the original 8B model with faster performance, while Schematron V2 Turbo focuses on maximizing throughput, achieving 4.14 requests per second, which is 2.5 times faster than its predecessor. Both models are available through a serverless API on Inference.net, with pricing designed to make web-scale extraction more accessible. Future developments include Schematron Pro, which aims to offer even higher accuracy without sacrificing throughput.

Catalyst is a newly launched platform designed to optimize production AI applications by integrating monitoring, evaluations, training, and deployment into a single system. It simplifies the process by using actual production data as a training environment, eliminating the need for synthetic environments and reducing costs significantly. Catalyst is compatible with existing OpenAI and Anthropic providers and can be integrated into a project with minimal code adjustments. The platform constructs training and evaluation datasets from real traffic, allowing for a continuous cycle of model improvement through its self-improvement flywheel, which manages data ingestion, training, and deployment. Catalyst's approach addresses the challenges of traditional reinforcement learning environments by providing a more efficient and direct method for improving production AI models. Currently available in public beta, Catalyst covers associated costs and offers users an opportunity to experience its full capabilities in optimizing AI systems using real-world data.

Inference platform is a public-beta full-stack system for improving production AI applications through integrated traffic monitoring, evaluation, supervised fine-tuning, and deployment. Designed specifically for deployed AI systems rather than general research, it captures real LLM requests through an inference gateway compatible with OpenAI- and Anthropic-style providers, then converts that traffic into training and evaluation datasets. Its workflow establishes baseline performance using LLM-as-a-judge evaluations, applies preconfigured training recipes and infrastructure, and deploys resulting specialized models to dedicated GPUs or customer-controlled environments. The company claims these models can match or exceed frontier-model quality at up to 95% lower cost, citing production uses in coding, extraction, calorie estimation, and research-agent applications. In contrast to reinforcement-learning approaches that rely on simulated environments and complex reward design, the platform argues that production data provides a more accurate basis for optimization, and it is offering free training and deployment during its beta period.