Laya

VeidtKV got Laya running on Radeon gfx1030 hardware, using a Hermes agent to do the porting. Code in Thotheris/laya-gfx1030.

FrontierMind tested the offline model on a MacBook Air at $0 per call on inbox triage, tagging 100 SAP decks, categorizing blog posts and a prefilter pattern.

20 real-world Chinese requests × 3 checkpoints, measuring accuracy, latency and whether confidence knows when it is wrong. Includes a Windows/Python 3.14 segfault workaround.

A Laya checkpoint trained on PolyAI's Banking77, the 77-intent banking support dataset, tagged for customer-support intent classification.

A survival arena game where every enemy's next move is a live typed question put to Jev or Laya, with probabilities coming back instead of text. Lets you watch the two decision models play head to head.

Fan-out SUMs, = NULL, UPDATE without WHERE, injection remnants and more, each with intent and schema context. Four typed questions per statement, identical payloads to both models.

Scores post style locally through a Laya-MLX helper as you scroll, using low-information and formulaic-rhetoric criteria and a conservative consensus score. Abstains when unsure.

Choice, Score and Noul over any System One backend with a confidence policy: if a backend errors or is unsure, the next is tried. Ships a TypeSafe-compatible server.

Soybelli connected Pac-Man to laya-mlx on Apple Silicon. The live view shows action probabilities, latency, tactical planning and safety interventions, with 0 output tokens.

Deterministic arcade physics, a browser flight recorder and a supervised Laya fine-tune that landed 30 of 30 held-out flights on Apple MLX. 58% less peak memory with 6-bit weights.

Leanest diffs a PR against main, asks one semantic judgment per spec and runs only the tests it cannot rule out. Andrea Bruno added Laya and classifier.dev as backends next to Jev.

Java port of Laya with choice/score/noul, a script-detecting router that picks the checkpoint, HuggingFace Hub fetching and bit-level parity with the Python ONNX output.

Tokenizer, mmBERT-base encoder and decision head in one binary for the 322M multilingual checkpoint. No Python, torch or BLAS; forward pass checked against a reference.

Side-by-side Snake: Laya local off a 1 GB footprint made 86.4 decisions a second (1,281 moves); Jev 1.13.0 over the API managed 3.2 (47 moves).

From the Prompt Engineer 48 video. Both models get the same input and typed questions; only the model differs. Laya needs no key.

AXERA-TECH's conversion of all three Laya checkpoints for its edge AI chips, reported at latency under 70 ms per decision on AX650/AX8850.

Google's LiteRT community conversion of the multilingual checkpoint (tflite), tagged for Android text classification. An English LiteRT build is published alongside it.

Native Swift library on Apple MLX. Call prepare once to download ~804 MiB of weights, then predict with choice questions straight from your app.

Local 15x15 Gomoku on Apple Silicon where each AI move is a single Laya choice among up to six candidate moves, with probabilities shown in the browser. The model was never trained on Gomoku.

An MCP server and HTTP sidecar for Laya adding a token-budget preflight that reports what would be truncated, a persisted calibration store, structured errors and a fix for a constant yes/no probability bug.

Homebrew menu bar server that turns n8n Text Classifier requests into Laya typed questions in ~40 ms. Weights ship inside the app; no network, unloads when idle.

One simulated Swedish villa, four controllers on the same minutes. Laya with a safety clip cut the bill from 985 to 530 SEK at p95 179 ms; a plain tariff EMS reached 385 SEK.

A Myanmar-language Laya checkpoint trained on Davlan/sib200, with a Gradio demo space from the same author.

An Arabic Laya checkpoint trained on AmazonScience/massive and facebook/xnli, with a companion laya-ara-rag ranking model from the same author.