Flide, a high-performance inference engine for Jev-style decision models, launches on Hacker News claiming up to 9.5x speedup over vLLM and SGLang. The open-source tool now supports over 30 popular local decision models, serving the Jev/v1/systemone API. Recent updates include new model families like Ainvest-Decision, Darwin-27B-RSI-Decision, and Know Line-31B-Gen1.5, alongside image-reading capabilities for the Know Line family.
The engine utilizes exact state forking technology to deliver faster, deterministic answers to structured questions about refunds, replacements, or escalations. Flide 0.1.0 initially supported Qwen3.5 and Gemma 4 chat models alongside One Jev and Open-Jev decision models. The engine serves requests through a standardized API endpoint, accepting state descriptions and questions with defined criteria and urgency.
Performance benchmarks across various models show mixed speed improvements, with some text-only models seeing 3.7x gains while models with image input show varying results. The project, documented on arXiv, positions itself as essential infrastructure for decision-making applications requiring rapid model responses.
Source: Hacker News · Summarized by HeadlinesBriefing