Releases·Model Releasefeatured
Developer releases Laya, a 421M-parameter open-source decision model trained with RLCD, with HF space, benchmarks, and repo
Convai Innovations released Laya, an Apache 2.0 non-autoregressive decision model built by Nandakishor M as an open alternative to TypeSafe's closed Jev API. It answers typed questions (choice, score, noul) in a single ~35 ms forward pass, with checkpoints on ModernBERT-large and mmBERT-base, and a community C++ implementation followed within days.
860 engagement4 sourcesFri, Sep 18, 2026, 06:25 AM
Developments
Community member lkarlslund released laya.cpp, an MIT-licensed C++ inference implementation with custom CUDA kernels and a JEV-compatible HTTP endpoint.
Mon, Sep 21, 2026, 02:33 AM
AI Overview
- Laya, a 421M-parameter non-autoregressive decision model, pairs a bidirectional ModernBERT-large encoder with a scratch Transformer head, resolving typed schemas in ~35 ms per forward pass
- The model trained on a single RTX 6000 Pro (96 GB VRAM) using a 100% human-annotated corpus of 25,000+ examples spanning intent routing, fact-checking, moderation, guardrails, rubric scoring, and multi-turn trajectories
- Training used RLCD, an unofficial policy-gradient reinforcement learning method optimizing decisions against strictly proper scoring rules for mathematically calibrated probabilities
- The model, runnable on low-end PCs, is available via an HF space, HF repo (convaiinnovations/laya), and GitHub repo (NandhaKishorM/laya)
- A community laya.cpp inference implementation built on ggml with custom CUDA kernels runs all three checkpoints without Python or PyTorch, hitting 366 questions/s BF16 at batch 1 on an RTX PRO 6000 (MIT-licensed)
- A 322M multilingual checkpoint built on mmBERT-base covers 100+ languages, with a built-in Router detecting input script and routing to the right checkpoint
- Author reports Laya beats TypeSafe's closed Jev API on AG News, emotion classification, and typed-decisions while running ~6–8× faster, from third-party Jev numbers rather than head-to-head runs
- Model card flags limits: base checkpoints near chance on typed-decisions without fine-tuning, Banking77 accuracy 0.425 vs Jev's 0.870 at 50+ options, overconfident probabilities requiring temperature fitting
Entities
Related
Releases·Tool Release
DaragonTech releases LibLayaX: run the Laya AI decision model in apps with no server or Python
A developer released four MIT-licensed open-source projects (LibLayaX, RLaya, LLaya, DLaya) that let applications call the Laya yes/no decision model directly on CPU or GPU with no server or Python. Built on Lars Karlslund's laya.cpp port of Nandakishor Mukkunnoth's Laya model, the code was written by Claude in about 32 hours.
A developer released four MIT-licensed open-source projects (LibLayaX, RLaya, LLaya, DLaya) that let applications call the Laya yes/no decision model directly on CPU or GPU with no server or Python. Built on Lars Karlslund's laya.cpp port of Nandakishor Mukkunnoth's Laya model, the code was written by Claude in about 32 hours.
9 engagement1 sourceredditSat, Oct 3, 2026, 12:09 AM
Social·Community
Laya zero-shot classifier matches its 78.33% MASSIVE benchmark, then loses to a keyword rule and TF-IDF on long texts
A developer reproduced the open-weights Laya zero-shot classifier's published MASSIVE (en) score of 78.33% to four decimals across three setups, then found it underperformed even a keyword rule and TF-IDF + logistic regression on their own document-classification tasks. The kit and write-up are public, and the model authors have been asked for comment.
A developer reproduced the open-weights Laya zero-shot classifier's published MASSIVE (en) score of 78.33% to four decimals across three setups, then found it underperformed even a keyword rule and TF-IDF + logistic regression on their own document-classification tasks. The kit and write-up are public, and the model authors have been asked for comment.
15 engagement1 sourceredditFri, Sep 25, 2026, 10:49 AM
Releases·Tool Release
stuntd: a local Jev-compatible server on Laya that learns from your own traffic (no API key needed)
A developer released stuntd, an Apache-2.0 local proxy for typed LLM decisions (choice, yes/no, score) that records traffic from a provider like the hosted Jev API, trains a small head on the frozen open-weights Laya encoder, and serves answers locally with fallback to the provider below a confidence threshold. It matters to practitioners because it cuts ~380 ms paid round trips down to ~22 ms local inference on closed-decision tasks, with self-promotion and demotion based on agreement rates.
A developer released stuntd, an Apache-2.0 local proxy for typed LLM decisions (choice, yes/no, score) that records traffic from a provider like the hosted Jev API, trains a small head on the frozen open-weights Laya encoder, and serves answers locally with fallback to the provider below a confidence threshold. It matters to practitioners because it cuts ~380 ms paid round trips down to ~22 ms local inference on closed-decision tasks, with self-promotion and demotion based on agreement rates.
33 engagement1 sourceredditWed, Sep 23, 2026, 02:13 AM
Social·Community
I literally built the Jev architecture one year back and completely open-sourced it with model, dataset and paper
A developer says he published the core idea behind the newly launched Jev architecture a year earlier — arXiv papers in March and September 2025 with open weights, datasets, and a pypi package — and criticizes the frontier lab's closed launch. Separately, a practitioner built a small classification head on the papers' ideas and reports beating Jev via API on a custom privacy-tiering task while flagging benchmark caveats.
A developer says he published the core idea behind the newly launched Jev architecture a year earlier — arXiv papers in March and September 2025 with open weights, datasets, and a pypi package — and criticizes the frontier lab's closed launch. Separately, a practitioner built a small classification head on the papers' ideas and reports beating Jev via API on a custom privacy-tiering task while flagging benchmark caveats.
Updated Tue, Sep 22, 2026, 03:56 AM — A follow-up benchmark post reports the fine-tuned head beating Jev's API 94.6% to 68.1% on a private-tier classification
2001.9k engagement3 sourcesreddit, hackernewsThu, Sep 17, 2026, 04:18 AM
Releases·Model Release
Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
During Jev launch week, Latent Space published an interview with Diogo Almeida, CEO of TypeSafe AI and InstructGPT co-author, on why he built a 'System One' model for production software rather than chat. Jev's launch video drew ~40M views, and its core training technique, Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for calibrated probabilities instead of human preference feedback. Practitioners get official patterns, cookbooks, and a coding-agents guide, plus use cases spanning computer use, games, linting, and analytics.
During Jev launch week, Latent Space published an interview with Diogo Almeida, CEO of TypeSafe AI and InstructGPT co-author, on why he built a 'System One' model for production software rather than chat. Jev's launch video drew ~40M views, and its core training technique, Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for calibrated probabilities instead of human preference feedback. Practitioners get official patterns, cookbooks, and a coding-agents guide, plus use cases spanning computer use, games, linting, and analytics.
0 engagement1 sourcerssMon, Sep 21, 2026, 10:13 PM