Laya is the first open model from Convai Innovations, a small AI company in Kasaragod, Kerala, India. It landed on Hugging Face on September 18, 2026, three days after TypeSafe AI opened early access to Jev. It does the same strange thing Jev does: it cannot write a single sentence. The difference is that you can download it, run it on your own machine and pay nothing.
Laya is a System One decision model. You give it a piece of state, such as a ticket, an email or a JSON document, plus a few typed questions. It returns one typed answer per question with a calibrated probability, in about 33 milliseconds, without generating any text. Convai’s tagline is “decisions, not text”, almost word for word Jev’s “Decisions, not strings.” That is no coincidence, and the story behind it is half the reason Laya went viral.
Key Features
- Typed Answers, Not Text: Laya supports three question types. Choice picks one option from a list you define. Score places the input on an ordered rubric, such as urgency from 0 to 3. Noul, Laya’s name for a boolean, returns a calibrated probability that something is true. You define the options at request time, so a new schema needs no retraining. The same caveat as Jev applies: “can’t hallucinate” means it cannot invent a category or break your JSON, not that it is never wrong.
- Calibrated Confidence: Laya is trained with what Convai calls RLCD, reinforcement learning against strictly proper scoring rules. In plain English, the model is only rewarded when its stated probability matches reality, so being honestly unsure beats being confidently wrong. Convai reports a calibration error of 0.081 against Jev’s 0.246. For automation, that number is the one that matters: it is what you gate escalation on.
- Speed: Convai measured a median of 32.8 milliseconds per routed question on a single Tesla T4 GPU, against 236 to 276 milliseconds for Jev’s hosted API. That is roughly 7.8x faster. In batches, it drops to about 7.2 milliseconds per question.
- Open Weights, Runs Anywhere: Code, weights, training method and benchmarks are all released under Apache 2.0. It runs fully offline, so the data never leaves your network. For healthcare, government or anyone under data sovereignty rules, that is the feature, not a footnote. The author says it runs on a low-end PC.
- Three Checkpoints and a Router: Laya is really a family. The English model is 421M parameters on ModernBERT-large. A 322M multilingual model on mmBERT-base covers 100+ languages. A third checkpoint is fine-tuned on Convai’s typed-decisions benchmark. A built-in router detects the script and picks the right model. That matters: the English model scored 0.000 on Khmer while reporting 0.952 confidence, the exact confidently-wrong failure the model exists to prevent.
Company Background
Convai Innovations was founded in 2021 in Kasaragod, Kerala, by CEO Nandakishor Mukkunnoth with co-founders Archana M and Anjali M. It started out building privacy-preserving and explainable AI for healthcare and edge computing. There is no $40 million seed round here. Laya was trained on a single GPU.
The founding story is a priority claim. In March 2025, Mukkunnoth published SalesRLAgent, a model trained with reinforcement learning to output calibrated probabilities instead of text. In September 2025 he followed it with a second paper on routing decisions by model confidence. Both were open, and both were pitched at a narrow sales use case that few people read. A year later, TypeSafe launched Jev with the same core idea, the same RLCD acronym and a lot of fanfare.
His reply was to make the idea general and give it away. “Not every AI problem requires an autoregressive chatbot,” he wrote. His launch post hit over 1,200 points on Hacker News, and the community split. Some argued that packaging, an API and a clean brand are the product, and that unread research does not count as shipping. Others were simply glad an open version exists. A fan even built a playground for it under the headline “Good research dies without marketing.” That line is the whole story in five words.
User Experience
- Getting Access: No waitlist, no API key. Install it with
pip install laya, or grab the weights from Hugging Face and the code from GitHub, which passed 9,000 stars. There is a live demo space if you want to try it before writing code. There is also no hosted API: you run it yourself, full stop. - The Developer Model: Same shape as Jev. You pass a state and a set of named questions, and you get typed answers with probabilities back. One practical tip: start the router with
Router(preload=True)to keep every checkpoint in memory. Otherwise, switching languages triggers a 7 to 10 second cold reload. - The Real Skill Is Fine-Tuning: This is the big difference from Jev. Out of the box, the base model scores 0.362 on Convai’s typed-decisions benchmark, barely above the 0.318 random baseline. The headline 0.766, which beats Jev’s 0.727, comes from the fine-tuned checkpoint. Convai’s own model card says to treat Laya as a fast base to specialise, not a zero-shot engine. You will also want to refit its confidence on your own data before trusting the numbers. Jev is a drop-in. Laya is a foundation.
- Where It Shines: Narrow yes/no safety calls. Convai reports 0.993 on spam filtering and 0.980 on phishing detection. LLM guardrails and jailbreak detection land around 0.76. Wider classification is weaker: 10-way support ticket routing scores 0.522 before fine-tuning.
- Known Friction Points: The benchmark caveat comes first, again. Every number is Convai’s own, and the Jev figures are third-party because Convai had no Jev API access. Second, choice questions degrade past about 20 options. On Banking77, a public test where the model must sort banking customer queries into 77 categories, Laya scored 0.425 against Jev’s 0.870. Third, you own the operations: the GPU, monitoring, and retraining as your data drifts. The top critical comment on Hacker News put it simply: requiring fine-tuning puts Laya in a different category from Jev.
Cost
Laya is free. There are no tiers, no seats, no meter and no API to bill you.
Self-Hosted (Open Source)
- Laya via Hugging Face and GitHub: Apache 2.0, commercial use allowed.
- Price: $0 per million tokens. You pay only for the CPU or GPU you run it on.
- Includes: three checkpoints, the language router, training code, benchmarks and a demo Space.
- Community ports: an unofficial MLX port runs it on Apple silicon Macs without PyTorch.
Jev already made decisions cheap at $0.042 per million input tokens, so price alone is not the story. The real argument is control: your data stays inside your boundary, and no vendor can change the terms. The cost does not disappear, it moves. You pay in engineering time for fine-tuning, hosting and monitoring instead of in API fees.
In summary, Laya is Jev’s open twin. Same idea, same typed answers, same calibrated confidence, but faster on its own benchmarks, free, and yours to run. The trade is simple: Jev works out of the box, Laya works after you train it. If you have a team that can fine-tune and host a model, Laya is one of the most useful releases of the year. If you want a drop-in classifier with no infrastructure, Jev is still the easier choice.
As with Jev, the vendor numbers deserve healthy scepticism until independent benchmarks arrive. A self-published speed comparison against a competitor you could not access is a claim, not a result.
But the bigger story is not which model wins. Within one week, a funded San Francisco lab and a small team in Kerala shipped the same idea. System One decision models are now a category, not a product. And for AI governance, the open version adds something Jev cannot: you can inspect the weights, retrain them on your own data and run them inside your own boundary. A calibrated confidence score lets you set the escalation threshold in code. Open weights let you audit the model that sets it. That is the combination regulated industries have been waiting for.