Decision AI Models: Jev, Fastino GLiDE & Open-Source Rivals
Explore the rise of TypeSafe Jev and Fastino GLiDE decision AI models, offering ultra-low latency, low cost, and fast parallel evaluation.
Stock photo for illustration only, not from the actual event
- Fastino Labs shipped 2 rival models while open-source developers published Jev reproductions
- TypeSafe Jev evaluates parallel questions without generating strings, preventing type errors
- Jev delivers 70 to 500 ms responses at $0.042 per million input tokens
- Vercel reported Jev as the fastest-adopted model in AI Gateway history
The artificial intelligence landscape is evolving rapidly after Fastino Labs shipped two rival models within a span of three weeks, alongside several open-source developers publishing Jev-style reproductions. This article explores how this model category operates, where it fits within the modern tech stack, and how various options compare against one another.
TypeSafe Jev accepts a state consisting of strings, arrays, or sets of name-value pairs, along with one or more questions. According to TypeSafe documentation, it supports three primitives where every question is evaluated in parallel and in isolation against the same state. Adding questions barely impacts response time, and because Jev never generates strings, TypeSafe states that it cannot return a type error.
Beneath the surface, TypeSafe utilizes a novel architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions, known as RLCD. While RLHF optimizes for human preferences, RLCD optimizes for calibrated probabilities where higher confidence directly correlates with higher accuracy.
Stock photo for illustration only, not from the actual event
Pricing serves as a crucial metric, with Jev costing $0.042 per million input tokens while output tokens remain free. OpenRouter lists a 32K context window, and TypeSafe reports end-to-end response times ranging from 70 to 500 milliseconds. TypeSafe built workflow evaluations spanning four distinct tasks: security incidents, agent trace observability, invoice processing, and customer service, using reference labels averaged from GPT-6 Astra and Claude Fable 5.1 at high thinking levels.
In terms of accuracy, Jev matched Sonnet 5 at a fraction of the cost and latency, though it still trails top frontier configurations by 6.3 points. Across individual tasks, Jev achieved 76.0% on customer service but only 61.8% on invoice processing. The practical rule of thumb is straightforward: if application code requires a bounded answer to branch on, a decision model is an ideal candidate, whereas human-readable outputs still necessitate standard LLMs.
The emergence of decision AI models marks a significant strategic pivot from open-ended text generation toward high-speed, cost-effective, deterministic logic processing. This shift empowers developers to integrate lightweight reasoning directly into automated software pipelines where millisecond-level response times and high reliability are paramount, effectively offloading specific structured tasks from resource-heavy frontier models.
Vercel reported that Jev became the fastest-adopted model in AI Gateway history, with nearly 13% of paid teams utilizing it within 24 hours—double the share of the GPT-5.6 family and over six times that of Fable 5.1. Furthermore, Simon Willison demonstrated fetching 100 candidates with BM25 and having Jev score each for relevance, successfully replacing an expensive LLM reranker with cheap, parallel score evaluations.
"Within 3 weeks, Fastino Labs shipped 2 rival models, and open-source developers published several Jev-style reproductions."
MarkTechPost
Arize and Langfuse have both introduced Jev-as-a-judge evaluators, with Langfuse labeling the feature as decision-model evaluators to accommodate future models, though its Jev support remains experimental. Additionally, a recent arXiv paper deployed Jev to interpret service contracts at the network edge, achieving a 22.4% reduction in median decision latency compared to DeepSeek and 61.9% versus Gemini at matched correctness levels.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment