Laya vs Jev: what the benchmarks actually report

Laya shipped six days after Jev, as open weights, and the first thing people did with it was race the two. A week in, the catalog has enough head-to-heads to say something careful. Every number below is as reported by its author; links go to the receipts.

Speed: not close, for a boring reason

• 86.4 decisions a second against 3.2, Laya local vs Jev 1.13.0 over the API, with the wifi turned off.
• 11x faster decisions on a 16 GB MacBook Air vs the cloud, on a Grok-built Tetris.
• 60 decisions a second from the MLX port under 1 GB of RAM; 3.7 ms per decision from the CoreML port on an M5 Pro.

The boring reason is the network. A Jev call pays a round-trip before it thinks; Laya's forward pass is 33 ms on a T4 and single-digit milliseconds on Apple silicon, and there is nothing to wait for. Any local model with sub-50 ms inference would win a decisions-per-second race against any API. The interesting number is not the ratio but that a 421M-parameter encoder is good enough to play the game at all.

Quality: the honest posts say Jev wins

Two of the cataloged benchmarks say so plainly. Tobias Wupperfeld's Snake benchmark ran Jev via API against Laya-MLX and Kev-4B and reports Jev "delivers the best quality no doubt." Tom Siwik's run has Jev winning "for a network-based model any day," with Laya-mlx tying his own JAX build locally, and argues the model matters less than how you define the decision space and batch.

That is roughly what you'd expect from a small open encoder against a closed model with a much larger training budget. Laya's own README does not claim otherwise on hard decisions; it claims speed, calibration and cost.

Cost: free vs metered

Laya is Apache 2.0 and runs where you already have a GPU or a Mac. Jev is metered. For a workload of a few thousand decisions a day the difference is pocket change either way; for a game loop, a browser agent, or a content firehose it is the difference between feasible and not.

What to take from week one

1. If the decision has to happen inside a tight loop (games, agents, UI), local Laya is the only option that fits the latency budget, and it is good enough for a surprising amount.
2. If the decision is hard and infrequent, the honest benchmarks say Jev is still more accurate. Read Laya's confidence and escalate low-probability cases.
3. Port quality is moving fast: MLX, CoreML, ONNX and GGUF all landed in the first four days. Check the catalog before you build your own.