Fine-tuned Laya vs Jev: Banking77 from 51.3% to 79.4%
Matched-budget fine-tunes on Banking77 and five more public tasks — prompt injection, SMS spam, emotion and request routing — with about 1,000 examples each.
BUILDS
Agents, routers, scorers and games — built on models that decide. Every entry links to the models it uses, so you can pull the same one and start from there.
113 builds · page 2 of 5
Matched-budget fine-tunes on Banking77 and five more public tasks — prompt injection, SMS spam, emotion and request routing — with about 1,000 examples each.
Collects vacancies from jobnet.dk and jobindex.dk, filters them against a candidate profile and ranks the matches locally.
A zero-dependency Rust runtime compiled to Wasm with WebGPU kernels; one import loads a pinned int8 model and nothing leaves the browser tab.
An Arabic checkpoint trained on MASSIVE and XNLI for language understanding, with a companion ranking model for retrieval.
Give it a numbered list of the page's controls and it picks the action and target — a drop-in for jev-ultrafast tooling that needs no screenshots.
A Myanmar-language checkpoint trained on SIB-200, published with a Gradio demo.
FluidInference's port, benchmarked at 3.7 ms per decision on an M5 Pro.
pythongiant's OpenCode plugin scores agent drift continuously with Laya and Jensen–Shannon divergence and charts it across the session.
One process serves the model API and an admin UI; agents ask for decisions over MCP, HTTP or the CLI and act on the top choice.
Four controllers ran the same minutes of a Swedish house. Laya with a safety clip cut the bill, but a tariff-aware energy manager did better; p95 was 179 ms per decision.
Anton Lobach ported the decision engine to Jolt, a Clojure implementation running on Chez Scheme.
A Homebrew-installed menu-bar app exposing an OpenAI-compatible endpoint, so n8n's Text Classifier gets Laya decisions in about 40 ms.
Dev Ed measured local Laya on a base M5 against Jev's hosted API on the same maze.
Adds a token-budget preflight that reports what would be truncated, a persistent calibration store and structured output to a warm MCP and HTTP server.
A local 15×15 Gomoku game: you play black, and every white move is one Laya choice among up to six candidate points.
A Swift library on Apple MLX with no Python or helper process: download the weights once, then call predict with choice questions.
Runs the checkpoint on Metal Performance Shaders at about 32 ms median and 2.1 GiB of memory on an M5 Pro.
Google's LiteRT community conversion of the multilingual checkpoint, for on-device text classification.
Leonardo Stenico moved the Needle extension from hosted Jev to local Laya, so finding passages by meaning needs no API key.
AXERA-TECH converted all three checkpoints for its AX650 and AX8850 chips.
An Agent Skill and MCP tool that gives Codex, Claude Code, OpenCode and Pi a local decision capability for routing work.
AmirTaha feeds the fruit, head and wall positions to the model and moves in the direction it returns.
mgear runs the ONNX weights in Unity through onnxruntime-unity and plays full games against the model.
The browser-use agent's decisions move from the hosted Jev API to laya-mlx, keeping the DOM snapshot, executor and safety checks.
Many of the first builds here were collected by madewithlaya.com. Every entry links to its creator’s own post, repository or site.