Open-Source Laya Challenges Jev in Fast AI Decisions

TL;DR: A Year of Research, A Sudden Spotlight

In September 2026, TypeSafe AI, a well-funded startup founded by OpenAI co-inventor Diogo Almeida, launched Jev, a non-autoregressive model for fast, structured decisions. The announcement generated significant hype. However, independent researcher Nandakishor Mukkunnoth claims he published the core concept a year earlier, in March 2025, and has now released Laya, an open-source alternative that he says is faster, more accurate, and fully transparent.

Laya is a family of models built on bidirectional encoders, designed for what Mukkunnoth calls "System 1" decisions—instant, reflex-like classifications that don't require generative text output. It runs in 32.8 milliseconds on a single GPU, supports over 100 languages, and is released under the permissive Apache 2.0 license. The project includes a PyPI package, Hugging Face checkpoints, and a live demo.

The Core Realization: System 1 vs. System 2

Modern AI pipelines often rely on large generative LLMs for simple tasks like routing tickets, flagging spam, or detecting prompt injection. This is inefficient, costly, and prone to hallucination. Mukkunnoth argues that these tasks need a different kind of model, one that mimics the brain's System 1: immediate, reflexive responses with honest probability scores.

"When an LLM outputs 'confidence: 0.95', it is just predicting tokens that sound confident," he writes. "There is zero mathematical calibration behind it." Laya aims to solve this by outputting only probabilities and numbers, making schema violations and hallucinations physically impossible.

Introducing Laya: Three Primitives, Three Checkpoints

Laya evaluates typed questions over any state (text, email, JSON) in a single forward pass. It uses three primitives: choice (pick one from a dictionary of criteria), score (place on an ordinal rubric), and noul (a boolean question returning calibrated P(true)).

Three specialized checkpoints are available under the main repository: laya (English, ModernBERT-large, 421M params), laya-multilingual (mmBERT-base, 322M, 100+ languages), and laya-typed-decisions (ModernBERT-large, 421M, optimized for agent observability and security alerts). The SDK supports selective subfolder downloads to save bandwidth.

The Multilingual Routing Problem

One of the most striking findings from Laya's 51-language sweep was that English models fail catastrophically on non-Latin scripts, often with high confidence. For example, ModernBERT-large scored 0.000 accuracy on Khmer while reporting 95% confidence. "The model's own confidence gives no warning when it cannot read the input script," Mukkunnoth warns.

To solve this, Laya includes a built-in Router that inspects Unicode scripts and stopword distributions in under 1 millisecond. It automatically routes text to the appropriate model, avoiding the 7–10 second cold-swap penalty by preloading checkpoints into memory.

Head-to-Head: Laya vs. TypeSafe Jev

Mukkunnoth provides extensive benchmarks comparing Laya (routed) to Jev 1.13.0, using third-party studies for Jev's numbers. Key results include:

• Accuracy: Laya outperforms Jev on AG News (0.950 vs 0.910) and DAIR Emotion (0.595 vs 0.480). On the typed-decisions benchmark, Laya scores 0.766 vs Jev's 0.727.
• Calibration: Laya's ECE is 0.081 vs Jev's 0.246—a 3x improvement in probability accuracy.
• Latency: Laya averages 32.8ms per question vs Jev's 236–276ms. On batched calls, Laya is 20x faster (7.2ms per question).
• Cost: Jev charges $0.042 per million tokens; Laya is free to self-host.
• Openness: Jev is closed-source; Laya is fully open with weights on Hugging Face.

The article also highlights 9 real-world workflows, including email spam filtering (0.993 accuracy) and phishing detection (0.980 accuracy).

Honest Limitations

Mukkunnoth is transparent about Laya's weaknesses. Choice questions degrade with more than 20 options—on Banking77 (77 labels), Laya scored 0.425 vs Jev's 0.870. Zero-shot performance is near random; the 0.766 score on typed-decisions comes from fine-tuning. Finally, base weights ship with raw temperature logits; fitting a scalar temperature per question type is recommended to reduce ECE from 0.466 to 0.081.

Why It Matters

The debate over who invented the non-autoregressive decision architecture is secondary to the broader point: not every AI task needs a generative chatbot. For high-volume classification, guardrails, and routing, a specialized, fast, and calibrated model is often more practical and cost-effective.

Laya's release challenges the proprietary approach of TypeSafe AI, offering a viable open-source alternative. As Mukkunnoth puts it, "It took a year of research... but the core realization remains: not every AI problem requires an autoregressive chatbot."