LAYA Q8 / LOCAL MODEL

Loading 524 MB browser model.

The first load downloads the Q8 encoder and head. Model files are stored in the browser Cache API when storage quota allows, so later loads can reuse them.

Laya is running locally…WASM inference in progress
LAYA LAB / OVERVIEW
Q8 · 524 MBWASM · CheckingCache · Checking
LAYA Q8 / IN-BROWSER DECISION INTELLIGENCE

NO TOKENS OUT.
JUST DECISIONS.

A 524 MB Q8 decision model running locally in the browser. No API key. No cloud inference. No generated text. Every interaction is a typed decision with probabilities you can inspect.

Built around what Laya is actually good at.High-signal classification, phishing/spam detection, news routing, confidence boundaries, and fast local decision experiments.

Decision Playground

Define any `choice`, `score`, or `noul` question and inspect the real distribution.

Inbox X-Ray

Spam, phishing, credential solicitation and pressure signals over one email.

News Pulse

A genuine 4-way topic decision: World, Sports, Business, or Science & Tech.

Prompt Firewall

An explicitly experimental view of jailbreak and suspicious-instruction signals.

Decision Race

Trigger a queue of events and watch them route into lanes one decision at a time.

Confidence Collision

Edit near-ambiguous text and watch decision margins move in real time on demand.

Local Privacy Vault

Run sensitive text locally and prove that inference sends zero input payloads to a server.

Observatory

Latency, confidence, tokens, option count, cache state and inference history.

ModelLaya Q8
RuntimeWASM
Weights~524 MB
Cloud inferenceNONE
Questions run0
01 / RAW PRIMITIVE

Decision Playground.

Use Laya directly. `choice` picks from named options, `score` returns an expectation over ordered levels, and `noul` returns P(true).

Question Builder
REAL Q8 INFERENCE
Decision Output
CALIBRATED DISTRIBUTION
StatusNOT RUN
02 / HIGH-SIGNAL SECURITY

Inbox X-Ray.

Paste any email. Four independent Laya questions reveal phishing, spam, credential solicitation and coercive urgency.

Incoming Email
USER TRIGGERED
Signal Map
4 REAL QUESTIONS
03 / 4-WAY CLASSIFICATION

News Pulse.

A visually clean 4-way topic decision aligned with the AG News-style benchmark: World, Sports, Business, Science & Technology.

Story
ONE CHOICE QUESTION
Newsroom Routing
FULL DISTRIBUTION
StatusNOT RUN
Topic probabilities
04 / PROMPT CLASSIFICATION

Prompt Firewall.

One genuine 4-way Laya choice classifies the prompt as normal, prompt injection, secret extraction, or social engineering.

This is still an experimental classifier, not a complete security control. The important change is that the categories now compete in one Laya choice distribution instead of four unrelated boolean scores.
Prompt
MANUAL TRIGGER
Prompt Classification
ONE 4-WAY CHOICE
StatusNOT RUN
Category probabilities
05 / EVENT-DRIVEN QUEUE

Decision Race.

Nothing runs by itself. Click Run Queue and Laya processes each item locally, one real inference at a time, while cards move into topic lanes.

Queue
USER TRIGGERED · NO AUTO LOOP
Ready.

World

Sports

Business

Science & Tech

Latency by item
Confidence by item
06 / DECISION BOUNDARIES

Confidence Collision.

Compare two similar texts under the exact same 4-way question. The interesting part is not only the winner; it is the margin between the top two classes.

A/B Decision Boundary
TWO MANUAL INFERENCES
A vs B distributions
07 / LOCAL-FIRST

Local Privacy Vault.

Run sensitive text through Laya locally. The page downloads model assets, but it never sends the user’s inference text to an inference API.

Private Text
LOCAL INPUT
Privacy Proof
RUNTIME TELEMETRY
P(true)
Inference API calls0
Input bytes uploaded0 B
RuntimeWASM
Model source
Threads
Latency
“0 input bytes uploaded” refers to inference payloads. Initial model/tokenizer assets are downloaded from Hugging Face/CDN.
08 / OBSERVABILITY

Decision Observatory.

Every real question run in this page is recorded locally: latency, confidence, token count, question type and option count.

QUESTIONS RUN

0

AVG LATENCY

AVG CONFIDENCE

Latency history
Question type mix
Confidence history
Recent inference log
#DemoTypeKTokensms