Laya-MLX ties a custom JAX build on a local game benchmark
Tom Siwik argues that framing the decision space and batching matter more than the model: over the network Jev wins, locally the two tie.
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.
35 builds · page 2 of 2
Tom Siwik argues that framing the decision space and batching matter more than the model: over the network Jev wins, locally the two tie.
Daniel Tremer's agent plays Hordes.io from screenshots: Apple Vision reads the screen, Qwen on MLX plans and Laya on Core ML makes each decision.
Akihito Koriyama's Semantic Browser follows HAL links by their ALPS semantics, with Laya-MLX giving each link a probability.
yankis's video runs Nimble, Decider, OpenJev, Laya and others through the same JevBench scoring; Laya finishes last.
Simplifying AI ran the same game and questions on both models; Laya made 86.5 decisions a second against the API's 3.2.
Tobias ran all three on a low-memory MacBook: Jev won on quality, the local models on speed and cost.
Sai Dutta's head-to-head: Jev leads on triage, guardrails, moderation and multilingual intent, while Laya wins on AG News.
Tony Dinh ran laya-mlx at about 84 ms per decision against Jev in real-time Tetris; Jev won all three games.
Daniel F's browser fighter has 21 characters, special moves, combos and replays, with the computer opponent run by Laya.
The announcement of the MLX port: under 1 GB of RAM, choosing the snake's next move sixty times a second.
Shantanu Goel trained Laya for about 24 hours on FreeDoom's first level from text-only observations; zero-shot Jev still played far better.
Many of the first builds here were collected by madewithlaya.com. Every entry links to its creator’s own post, repository or site.