Hosted Laya: $1 free, no card
laya.tools
Submit

Seen on X

What people are building with Laya.

Demos, games, ports and benchmarks posted on X, picked by hand. Each one links back to the original post and, where there is one, to the project in the directory.

Deepesh Bansal

@deepeshbansal_ · Sep 30

An audio-native Laya classifier answers typed questions without transcription

I trained an audio-native, Jev-like classifier on Laya Gemma 4 audio encoder + a 34M projector, so Laya reads audio like text. No transcription and no generation, just typed answers in one pass. vs Whisper-large-v3 → Laya: • Better on yes/no + noul questions • 2.5–8.6× faster • Same WER on phone calls Built for routing, evals, and sitting next to a live voice model with tools. Write-up: #genai #voiceai #jev #laya #google #gemma #gemma4

♥ 5↻ 069 viewsView on X

Rok Benko

@rokbenko · Sep 29

quackd coordinates robot fleets with Laya for multiple-choice decisions

I started building an open-source version of this AI <> Microduck use case on August 28, before GPT-6 Astra even came out. quackd is an open-source multi-robot CLI that lets you connect all your robots, command them, and let them work together, each with an LLM for a brain (Claude, OpenAI, Gemini, Grok, or local via Ollama or vLLM) and a decision LLM for multiple choice (Jev, Laya, Kev). Drives Microduck, Open Duck Mini, LeRobot, XLeRobot, AlohaMini, ToddlerBot, or any ROS 2 base from a laptop, offboard.

♥ 5↻ 0164 viewsView on X
Project: quackd→

Boundless CN

@boundless_cn · Oct 2

Laya joins Vercel AI Gateway for typed decisions in agent workflows

Laya 现在通过 Boundless 上线 Vercel AI Gateway：从「生成答案」到「直接做决定」 继 DeepSeek、GLM 之后，Laya 是一个很好的补充 它不负责聊天，而是专门做判断的 open-weight model：分流工单、筛销售线索、排反馈优先级，适合 agent 和各种 workflow 越来越多模型正在通过 Boundless 进入 Vercel 的开发者生态

♥ 5↻ 0426 viewsView on X

Cryptofic

@MyCryptofic · Oct 6

Acurast runs Laya demos for games, moderation, and inbox sorting on phones

I just tried the Laya Demos by @Acurast, and it's mindblowing to see what would AI and decentralized computing looks like in the real world. 3 mini-sections to explain why this is huge: 1. What's Laya and how does it work? 2. What's $ACU role in all of that? 3. What's the impact on ACU token price? Let's dive in. 1. What's Laya and how does it work? Laya - built by ConvAI Innovations and run by Acurast - is an open-weights decision model, not a chatbot. How things work with Laya: Laya picks one and returns probabilities. No essay, no chat completion. About 0.2-1 second on CPU only. You can see the loop demo live, actually: Snake, Tetris, Doom, mail sort (inbox / spam / phishing), newsroom labels, a prompt-injection guard, chat moderation. No signup required. Every answer comes from a real Android phone on the Acurast network, with country, chip, decisions served, and an on-chain link. A decision service that used to imply a data center, fits on one phone, somewhere in the world. 2. What's ACU role in all of that? Essentially, ACU is how that gets deployed. Scheduling a job pays network fees in ACU. Here's how the economics work with Laya deployments: - Acurast’s very own Cargo: Laya template is under 1 ACU a day for a private instance. - Processors (your smartphones and mine!) earn from inflation (5% a year: 70% staked compute, 10% benchmark rewards) plus a busy bonus when a phone is actually executing. - ACU token holders can delegate stake without hardware. 3. What's the impact on ACU token price? Long term, price only follows if demos become real-life, recurring paid deployments: classifiers, guards, agent backends. - Try it: - Become a compute provider by turning your unused phones into processors and generate ACU in the process:

♥ 5↻ 0138 viewsView on X

Ruban

@RubanBhatia · Oct 12

laya-guard checks coding-agent shell commands locally before execution

A 322M parameter model on your CPU can now judge every shell command your coding agent runs, in about 90 ms. I like where this is going. I'm not sold on calling it a security boundary yet. laya-guard sits between the agent and the shell and answers allow, review or block before anything executes. No API call, nothing leaves your machine. It has adapters for Claude Code, Codex CLI, Cursor, Gemini CLI and more. The numbers are decent. Self-reported, it blocked 25 of 27 AgentDojo attacks and kept 92 of 97 benign tasks. At 90 ms you can put it in the hot path without feeling it. Two things keep me cautious. The hook fails open by default, so if the judge breaks, commands just run. A guard that disappears when it crashes is a speed bump. And it judges each command alone. A commenter pointed out that npm test sits on the safe list, but the agent can edit package scripts first and then run it. Write a bad script in step one, run bash in step two, and each step looks harmless. So the check needs repo context, or a human on the risky edge. Cheap local judges are great for catching the obvious stuff. Would you run yours failing open or failing closed?

♥ 5↻ 0129 viewsView on X

Mooncast Productions

@MooncastOnline · Oct 14

A llama.cpp benchmark tests Laya on toxicity and shell-command safety

I started messing with Jev-style classifiers last night. Super easy to get running with llama.cpp! So I put together a simple decider benchmark that checks for toxicity in chat and also looks at a bash command and determines its safety. Laya is ~800Mb on disk and managed about 65% accuracy. Clef is ~15Gb and managed 96%. Still playing around, but I can see lots of applications.

♥ 5↻ 0203 viewsView on X

Alok

@analogalok · Oct 28

Qwen 3.5 0.8B decision model benchmarked against Laya

my 0.8B decision model wasn't bad at WinoGrande and ANLI. it was biased. i looked at what it actually answered, not just the accuracy: it picked the second option 2 out of 3 times, and almost never answered "contradiction", even though a third of the answers were. fixed it with 2,400 class balanced rows. Qwen 3.5 0.8B AA Decision: WinoGrande 50.0 -> 55.0 ANLI 35.5 -> 41.2 BANKING77 56.1 -> 54.2 (the cost) vs Laya, Q4_K_M against their Q8_0: ARC-C - 58.7 vs 31.1 ✅ ARC-E       - 77.5 vs 47.9 ✅ BANKING77   - 54.2 vs 33.0 ✅ WinoGrande  - 55.0 vs 52.1 ✅ ANLI        - 41.2 vs 65.4 ❌ 4/5 now. ANLI still loses. updated GGUF, same repo: /v1/systemone, one forward pass, mainline llama.cpp.

♥ 5↻ 0535 viewsView on X

Helder Nakaya

@helder_nakaya · Nov 2

Laya-Med reaches 85.6% diagnostic accuracy at 64 times the speed

LLMs are powerful, but they are also slow and expensive to run. Models like Jev and Laya take a different approach. They are fast, lightweight, and can assign probabilities to predefined outcomes directly from free text. In our study, Laya-Med achieved 85.6% diagnostic accuracy, compared with 83.1% for Google's MedGemma 27B, while running about 64 times faster! And here's the exciting part: inference is so fast that we can build dashboards where the probabilities of different diagnoses change in real time as a physician types clinical notes. Not every medical AI application needs a massive LLM!

♥ 5↻ 01.3k viewsView on X

0xAI42.exe

@0xAI42exe · Sep 28

laya-mlx plays Snake at 60 decisions per second on an M3 Max

Introducing a version 50 times faster than Jev, running on your device: laya-mlx! Occupies a maximum of 1G memory only on your device Laya is an open-source classification system similar to Jev, based on text output probabilities I ported it to MLX and made some performance optimizations! In the video, this model is playing Snake on my local M3 Max This model can play Snake at a decision speed of 60 times per second!

♥ 4↻ 0174 viewsView on X
Project: laya-mlx→

Qtum.ai

@QtumRouter · Sep 30

Qtum AI Router hosts Laya models for structured decisions

Laya is live on Qtum AI Router. Three decision models. You send a ticket, email, or JSON plus typed questions. You get a structured answer, not a paragraph. • laya — English, 512 tokens • laya-multilingual 100+ languages, 1024 tokens • laya-typed-decisions specialist workflows, 1024 tokens 21 credits / 1M input tokens. Output is free. Same key, same balance, hosted on Qtum hardware. Use it for triage, tool routing, urgency scoring, and model-tier gates. Let System 1 decide. Keep the big model for the work that needs it.

♥ 4↻ 276 viewsView on X

Posted a Laya demo on X? Mention “Laya” and it gets picked up automatically, or submit the project.