# What Is Laya? Laya vs Jev, Explained Simply > Source: > Published: 2026-09-23 14:01:04+00:00 If you’ve ever wired an LLM into production, you know this exact headache. You take an incoming support ticket, paste it into a prompt, and beg the model for clean JSON: return valid JSON, no markdown backticks, never apologize. And then 2 a.m. rolls around. The model apologizes anyway. Or it wraps the JSON in backticks, or invents a routing department missing from your database. Your parser crashes, PagerDuty fires, and you’re writing ugly regex in the dark while adding retry loops with exponential backoff. Soon you’re chaining a second model just to police the first. It’s brittle, expensive, and completely ridiculous. We’ve been using storytelling machines to make operational judgment calls. That’s the mistake. Think about a doctor looking at an X-ray. You don’t ask for an essay on bone biology. You just ask: is it broken, yes or no, and how sure are you? That’s the entire premise of a decision model. Instead of slowly generating sentences token by token, it looks at your input text and returns an answer in one of three clean shapes: Choice, Score, or Noul. A Choice question picks one label from a list you provide, like routing an email to billing, support, or sales. A Score question rates text on a scale, like customer frustration from one to five. A Noul question asks a straight yes-or-no question and returns an exact probability. And these models are actually calibrated. That means when the system says it’s 80% confident across a hundred tickets, it’s right about eighty times. It’s a group track record across real traffic, not an empty promise on a single guess. The architecture is almost embarrassingly simple. You don’t ask a chatbot to write a paragraph and fish out a label with string matching. You give a classifier a typed question and get back a probability. You might ask: that’s it? Yes. That’s the point. That simplicity is why it works. There’s no prose to parse, no way to answer outside your allowed options, and the whole decision happens in a single forward pass instead of a grinding generation loop. On September 15, 2026, TypeSafe AI launched Jev, as detailed in [Maxim’s technical breakdown](https://www.getmaxim.ai/articles/what-is-jev-system-one-model/). The company was founded by CEO Diogo Almeida, a former OpenAI researcher, and backed by $40 million in venture funding. Jev is a proprietary cloud API with closed weights. TypeSafe AI charges $0.042 per million input tokens, while output answers are completely free. Calls clock in between 70 and 500 milliseconds. TypeSafe AI claims Jev runs 193.6 times faster and 444.6 times cheaper than frontier chat models on comparable evaluation tasks. It also packs a massive 64,000-token context window to evaluate whole documents at once. Shortly after Jev debuted, developer Nandakishor M and Convai Innovations released an open-source alternative named Laya, hosting the [model card on Hugging Face](https://huggingface.co/convaiinnovations/laya). It’s released under the Apache-2.0 license, meaning anyone can download it and self-host it for free. Laya glues a ModernBERT-large backbone to a small decision head, landing at 421 million parameters. It evaluates the exact same three question types as Jev. It was trained using reinforcement learning with proper scoring rules, a technique called RLCD where the model gets rewarded purely for outputting honest, calibrated probabilities. Because Laya is compact, a single question runs in 32.8 to 39.5 milliseconds on an inexpensive Tesla T4. The first time you see a probability land in thirty-three milliseconds without a single line of parsing regex, it feels like cheating. The project also ships a 322-million parameter multilingual model based on mmBERT covering over a hundred languages. This wasn’t built overnight. In March 2025, the author published the underlying approach in an [arXiv preprint](https://arxiv.org/abs/2503.23303) to predict sales conversion probabilities during live customer calls. When the author posted the story on Hacker News as [“I built non-autoregressive decision models with RL a year ago”](https://news.ycombinator.com/item?id=49765348), it racked up roughly 1,100 points. The longer version of that story is on [Dev.to](https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me). FeatureLayaJevMakerConvai InnovationsTypeSafe AIModel WeightsOpen (Apache-2.0 on Hugging Face)Closed (Cloud API only)CostFree self-hosted compute$0.042 per million input tokensLatency (single question)32.8–39.5 ms on Tesla T470–500 ms via APIAccuracy (typed-decisions)0.7660.727Calibration Error (ECE)0.0810.144Context Length512 tokens (1,024 multilingual)64,000 tokens Now let’s be honest about where Laya falls flat on its face. Laya isn’t an oracle out of the box. Its raw base checkpoint scores a painful 0.362 zero-shot accuracy on the typed-decisions benchmark, which is essentially random guessing. That shiny 0.766 score belongs to a checkpoint fine-tuned on that benchmark’s own training split. Next, Laya chokes on too many options. On Banking77 with seventy-seven labels, Jev hits 0.870 while Laya crumbles to 0.425. Laya allocates a fixed token budget for choices, leaving only three or four tokens per label when facing dozens of choices. Laya also ships overconfident. Its raw calibration error sits at 0.213. You must run temperature scaling on your own validation data to coax it down to that crisp 0.081 score. Finally, Laya limits English text to 512 tokens, so long documents must be chunked manually. Both models share common blind spots: they process text only, cannot count or do arithmetic reliably, and adversarial prompt text can steer their scores. Deciding what to run depends on your engineering constraints. Take Laya if you care about data privacy, want zero per-token bills, and run your own GPUs. It’s fantastic for short payloads like support tickets, assuming you can fine-tune on your company’s historical labels. Pick Jev if you’re chewing through long documents, need to categorize across dozens of labels without fine-tuning, and would rather swipe a credit card than keep GPU servers alive. We spent years treating language models like digital humans who must converse before they act. But when you just need code to branch reliably, conversational chat is pure overhead. Laya puts the weights directly in your hands. You don’t need permission from an API provider or a massive cluster. You can spin it up on a cheap GPU, send it a question, and see the simple idea work for yourself. [What Is Laya? Laya vs Jev, Explained Simply](https://pub.towardsai.net/what-is-laya-laya-vs-jev-explained-simply-f7125dd3e582) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.