Laya

One import, no server. Dependency-free Rust compiled to Wasm with WebGPU kernels; pulls a pinned int8 pack from Hugging Face and keeps it in the browser. Live Tetris included.

Collects vacancies, filters them against a candidate profile, ranks with the multilingual Laya checkpoint and prints IT and warehouse tables plus a Markdown report.

Matched-budget fine-tunes on Banking77 and five more public workflows (prompt injection, SMS spam, emotion, request routing), plus two Aegis chat-moderation fine-tunes.

Automotive, embedded, enterprise software and product decisions, three trials each, with questions, expected answers, raw responses and a static HTML report.

Self-hosted console that runs VirusTotal lookups through System One models and evaluates them against each other on a labelled dataset. Prompts and criteria are versioned JSON.

Reproducible zero-shot benchmark on informal Moroccan Darija reviews in Arabic script and Arabizi, positive/neutral/negative. Deterministic seed-42 split, no training.

Paste a nervous message and watch tone (choice), formality 1–5 (score) and fight risk (noul) update as you type, 40–110 ms per call. Misses subtle passive-aggression.

Anthony Alcaraz's talk demo routes typed decisions through Jev and shows Laya's answers beside them. On the TSMC run Laya rejected every true claim; that's the point.

A local MLX Laya run against Jev on 40 real support tickets in EN and Thai. Laya was faster but not close on accuracy yet, per the author.

A frozen-protocol comparison of Jev 1.13, PrismNLI-0.4B and Laya with raw predictions, calibration and latency published. Laya ties Jev on the emotion dataset and trails it by 10 to 33 points elsewhere.

Seven zero-shot extraction and classification families demoed, benchmarked and cross-compared, with sizes, licences and costs in one table.

TypeScript library where one set of question definitions serves a zero-dependency Promise client and an Effect service. Never invents a probability it was not given.

SwiftPM package ported from laya-coreml. Point it at the aac6fef Core ML bundles (general or ANE) and call predict with choice, score or noul; no Python at runtime.

Browser runtime for Laya decisions on ONNX Runtime Web. WebGPU with SIMD Wasm fallback, Web Worker friendly, CPU embedding slicing to get past storage-buffer limits.

Live at laya-web.pages.dev. 524 MB of quantized English weights cached in the tab, onnxruntime-web over WebAssembly, and a parity page checked against the PyTorch reference.

Replayable grid sim with moving workers and forklifts. Laya answers action, collision_risk, path_blocked and needs_operator in one call; collision prevention stays deterministic.

The game enumerates placements with their holes, heights and clears; Laya picks one from text. The teacher is a board-evaluation formula, not Jev. Walkthrough video in Japanese.

Brain function collapse's live site where Laya plays Tetris, Flappy and a lane runner on a laptop, about 200x faster than a hosted LLM. Backed by the laya-playground repo.

Will Sargent's stdio MCP server on fastmcp. Classify text, score against a rubric or answer yes/no from Claude Code or any MCP client without sending input off the machine.

GGUF conversions of both the English and multilingual Laya checkpoints for llama.cpp-style local runtimes.

Postgres FTS, pgvector and RRF first stage; Jev, Voyage, GPT-4.1-mini, GPT-5.6-luna and open-weight Laya as rerankers on two Finnish legal corpora. Recall, MRR, nDCG, latency.

A companion service for Lidarr that reads the manual-import queue, scores every candidate release and track in one pass with Laya, and picks the best match. Runs locally at no cost.

A vision variant of Laya on Hugging Face: the same choice/score/noul question shape, with an image as the state.

MLX port of Laya with performance work. Under 1 GB of memory on-device; the demo plays Snake at 60 decisions per second on an M3 Max.