Tag
作者以语音 AI 编排平台产品负责人的视角，分析 Jev 及同类分类模型在电话语音机器人中的实际用途，并梳理了 Laya、Clef、CMF、Nimble、Kev 等竞品生态。文章指出单一 LLM 无法独立支撑生产级对话，分类模型在延迟、成本和结构化输出上仍不可或缺。
TypeSafe 发布 Jev 1.14.0 模型更新，并通过 "danger zone" 实验性功能引入图像支持、新的 box 边界框类型以及更长的上下文长度，面向开发者与创意用户开放预览。
Cloudflare 发布了新的多模态模型 Clef-Omni，这是一个基于 Qwen3-Omni-30B-A3B-Instruct 后训练的 30B-A3B 混合专家多模态模型，可将状态和结构化问题模式一次性转换为概率决策，无需自由文本生成。
Microsoft 推出 Decision-1 决策打分模型，用于代理与工作流的路由、分类、验证和控制，单次推理即可输出各选项的校准概率。该模型基于 Qwen3.5-9B 后训练，据称中位延迟比 GPT-6 Sol 快约 35 倍，现已在 Microsoft Foundry 上线。
本文提出 FinVector-Market-4B，使用 rank-16 LoRA 在 22,000 条语料上对 Qwen3.5-4B 进行结构化金融任务的适配研究，表明在匹配提示契约下，紧凑型领域适配能带来超出输出格式学习的显著任务收益。
Jev-Mem:一篇来自UT Dallas的论文提出用TypeSafe AI的Jev结构化输出模型替代LLM来处理智能体记忆系统中的大量小型决策（如记忆分类、关联判断、检索停止条件），仅在最终生成答案时才调用LLM，从而显著降低token开销。
Simon Willison released llm-openai-decisions 0.1a0, an LLM plugin for OpenAI's new Decisions API, which supports predicate/choice/score question types and, via the gpt-6-luna model, image input alongside text.
Kolibri-1, an open-weights model from Aleph Alpha, is shown playing Breakout without fine-tuning, using optimized inference that turns action probabilities into moves with under 25ms latency per decision.
TypeLLM, an alternative to Jev for LLM structured output, supports conditional fields where earlier answers determine which fields are generated next (e.g., extracting an amount only for invoices) — all within a single API call using `when`.
Cloudflare released Clef, an open-source (Apache 2.0) 27B multimodal decision model on Hugging Face that converts a state and a schema of typed questions into probabilities for all options in a single forward pass. Post-trained from Qwen3.8-27B, Clef has no free-form text generation and is API-compatible with Jev and SystemOne, with a smaller Clef-Flash variant available.
Cloudflare released Clef, an open-weights 27B multimodal decision model that takes a state and a schema of typed questions as input and returns probabilities for each option in a single forward pass, with no free-form generation or output parsing. It is post-trained from Qwen3.8-27B, ships on Hugging Face with a smaller Clef-Flash variant, and is compatible with the Jev/SystemOne API.
Cloudflare released Clef-Flash, a 9B multimodal decision model built on Qwen3.5-9B that converts a state plus a schema of typed questions into per-option probabilities in a single forward pass, with no free-form text generation or output parsing.
The article discusses a benchmark where coding agents reconstructed a scientific flow diagram as editable PowerPoint slides, arguing that editable artifacts better test visual understanding by revealing structural comprehension versus pixel-level reproduction.
The article highlights that most LLM speed-vs-accuracy comparisons overlook the trustworthiness of structured answers, and showcases JEV as a tool that ensures consistent decision outputs through its structured interface.
Jev can generate game levels in real time, offering faster and cheaper structured output that could significantly impact game development.
TypeSafe AI released Jev, a semantic decision engine designed for fast, low-cost AI decisions in software systems, avoiding the inefficiencies of generative LLMs for simple choices.
Building your own AI agent toolkit isn't complex; the minimal version includes tool schema validation, structured output parsing, and failure retries, but its true value lies in debuggability, allowing you to pinpoint issues through logs.
This guide details a fine-tuning recipe using Group Relative Policy Optimization (GRPO) with the TRL library to enhance the LFM2.5-350M model's structured output compliance, improving IFStruct benchmark performance from 22.6% to 29.7%.
This paper introduces the trie automaton, a specialized constrained decoding mechanism for finite sets that precomputes token masks via Aho-Corasick matching, achieving up to 29x end-to-end throughput improvements over XGrammar in vLLM batch serving while guaranteeing 100% output validity.
The author presents a method using GBNF grammars in llama.cpp to constrain local AI model outputs, ensuring valid JSON. A compiler is built to convert tool schemas into grammar rules, with per-turn narrowing for relevant tools.