What is Jev? (and is laya better?)

Not the same stack. laya runs locally through FluidInference's CoreML build, kev runs locally, and Jev is called through the hosted API, so hardware, batching, quantisation and network all differ between contenders. The page documents no matched settings, and the entry flags this as one informal comparison from GitButler's channel with a product promo link attached.

A Tetris match between three models and a GitHub Settings filter. No win counts, no latency percentiles, no error tallies and no repeat runs appear in the published text, so you cannot tell whether any difference would survive re-running. Both tasks are also low-stakes with clean labels, which flatters every model involved.

A public test repository is linked, which helps, but the page documents no pinned Jev version, no laya or kev checkpoint, no prompts and no harness config. Two of the three models also shift between runs as upstream repos change. Treat the repo as a starting point for your own comparison, not as a reproduction script.

The page never defines better. The only dimensions named anywhere are who won the game and how the settings filter came out; no cost or latency figures are recorded. Running laya locally swaps hosted per-token pricing for hardware you already own plus your own maintenance, so the comparison only means something against your real usage pattern.

Scott Chacon on GitButler's channel, with a product promo link attached. The entry flags the results as one informal comparison rather than a benchmark.

Through FluidInference's CoreML build, while kev runs locally and Jev is called through the hosted API. Hardware, batching, quantisation and network all differ across the three.

Three. The two tasks are a Tetris match between three models and a GitHub Settings filter, and the page records no win counts or repeat runs.