Laya · 18 pages · read 2026-09-24
Every Laya guide, in one place
18 pages about an open-weights model that is nine days old, grouped by the question each one answers. Two of them are ours end to end — the probe we ran against the released weights, and the same thirteen cases sent to both Laya and Jev — and the rest say what other people published, with the source and the date on the page.
Weights read 2026-09-24 Counts read from the same files the pages render New here? What Laya is, then how it behaves next to Jev.
What Laya is, and what it costs to run it
Five pages, in the order the questions arrive: what it is, what is in the box, whether it does what Jev does, and what "open" actually covers.
-
Laya, the open-weights decision model
The front door: what shipped on 18 September 2026, and what our own run of the weights showed.
-
Laya, specified
The long version — checkpoints, limits, the probe we ran, and the numbers we could not reproduce.
-
Laya vs Jev, on the same thirteen cases
Both models, one harness, the same adversarial set — where they agree, where they diverge, and what divergence costs.
-
Self-hosting Laya
CPU speed, token budgets and two sharp edges we hit while running it locally.
-
What Apache-2.0 covers
The licence, the three weight sets, and the part of a hosted model that no licence can hand you.
How to build with it
Laya's first week produced runtimes rather than applications, so this group is about getting it running rather than about product patterns.
-
Laya with MLX
The port that travelled: what changes on Apple silicon, and the two places the runtime stops behaving like the reference server.
-
The model card, the weights and the reading
Every link worth having, including the two hosted demos and the four ports still being argued about.
15 records
What people built with it
Ports, servers, agents and games. The same records appear under a kind of work as well as under a column — that is the second axis, not a second copy.
-
Every record, by kind of work
Six cuts across the whole corpus: ports, servers, agent tooling, browser decisions, games, benchmarks.
-
Built with Laya
Each project with a dossier: what it does, how it works, what we checked.
19 records
-
Runtime ports
Getting the weights onto someone else's hardware: MLX on a Mac, Core ML on the Neural Engine, ggml in C++, candle in Rust, ONNX in Node.
24 records
-
Serving and API
Standing the model up as something else can call, usually behind Jev's own wire format.
10 records
-
Agents and tools
Typed decisions inside an agent loop: routing, evals, guardrails, memory, and the plumbing around them.
7 records
-
Browser and computer use
Deciding the next click, key or field on a real screen.
2 records
-
Games and demos
The model playing something you can watch. Most of this model's traffic arrived through one of these.
10 records
-
Benchmarks and evaluation
Numbers somebody actually ran — including the ones that say Laya lost.
24 records
Where the records are
The columns, with the counts they carry today.
-
Laya on X
Posts read in full and described in a sentence of ours.
28 records
-
Laya videos
Local installs, demos and head-to-head games, with durations read from the platform.
11 records
-
Laya on Hacker News
The author's own submission and the threads it started.
6 records
What this page is not
Not a catalogue: the columns below hold the records themselves — 18 shelves and cuts, with their live counts. This page is the reading, and it grows by one line each time we write another one.